Technical leadership standards
Criteria verified as of August 2026. Re-verify on the web before committing to anything (§8).
1. Scope and triggers
Covers the decisions taken by whoever leads technically and the artifacts in which they are
written down: career tracks and the real scope of each level, staff+ role archetypes, who decides what
and how fast, the decision record, the design document and the design review, technical debt
as a budget line, the individual conversation and performance evaluation, team
sizing and its dependencies, and the resolution of technical disagreements.
Triggers: "career ladder", "level ladder", "senior/staff/principal", "IC vs. management",
"staff archetype", "tech lead", "ADR", "decision record", "one-way door",
"reversible", "design document", "design doc", "design review", "technical RFC", "technical
debt", "refactor budget", "blameless postmortem", "1:1", "feedback", "performance
review", "individual productivity metrics", "lines of code", "commits per person",
"DORA per engineer", "SPACE", "cognitive load", "Team Topologies", "team size",
"inter-team dependencies", "technical disagreement", "disagree and commit", "responsibility without
authority".
Guiding principle: technical leadership is a set of decisions and artifacts, not a character
trait. It is not evaluated by how the person is perceived, but by what exists in writing and by
what the team achieves: decisions recorded with their reversibility, designs reviewed before
building, interfaces between teams defined, debt with an assigned budget and people with
specific, dated feedback. A falsifiable test applicable to any claim about leadership:
name the artifact it produces, who reads it and which decision it changes. If it cannot be named, the
claim is motivational and does not get written here.
Hard corollary: "lead by example", "foster trust" and phrases of the same genre fix
nothing. They are not criteria, they are decoration. Their actionable equivalent always exists and is the only thing
this document accepts: "every architecture decision lives in an ADR with declared
reversibility", "the postmortem names no people and the leader defends it to management in writing".
Not applicable:
technical-hiring-standards: reciprocal and strict. Here you decide what profile is needed,
why, and which gap in the team it fills (§3.7); there how a candidate is measured — rubric,
exercise format, method validity, bias, legal regime and onboarding. One-sentence boundary:
leadership defines the role; the selection process is the measuring instrument.
project-management-standards (already written): delivery, the plan, the commitment, the
estimate, the risk and the stakeholders are theirs, including the record of management
decisions. Here the technical decision and team development. Boundary: if the question
is "when will it be ready and with what risk?", it is theirs; if it is "how is it built and who decides it?",
it is ours. The decision record format is the same in both (§3.3) so as not to fork
the repository.
product-discovery-standards: what gets built and why — problem,
user, hypothesis, prioritisation by value and the decision to kill an idea. Here the what is not decided.
enterprise-architecture-standards: architecture governance for the
organisation, capability roadmap and application portfolio. Here the scope is the team
and its immediate neighbourhood, not the company. Reciprocal declared in their §1.
software-architecture-patterns-standards and refactoring-tech-debt-standards (already written):
the technical criteria are theirs — which pattern applies, when a module is refactored
and with which technique. Here only the decision to invest in it, its budget and how it is defended
(§3.5). Both already declare the reciprocal in their §1.
microservices-architecture-standards: service boundaries and concrete contracts. Here, the
organisational reflection of those boundaries (§3.6).
code-review-standards (already written): the review criteria are theirs — what blocks, how
the comment is written, what a high-risk change requires. Here only the rule that review
is not used as an instrument of power (§3.8) and who arbitrates the disagreement that review
does not close.
testing-qa-standards: test strategy and gates. Here, the decision to fund it.
sre-practice-standards: SLOs, error budget, on-call, capacity and the DORA metrics are
theirs. Here, solely, that DORA is a system measure and its individual use is forbidden
(§3.9, §7).
incident-management-standards: the incident process and the blameless postmortem are theirs
— roles, timeline, actions. Here what corresponds to leadership and only that: protecting the
blameless postmortem when there is pressure from above to name a culprit (§3.10).
platform-engineering-standards (already written): the platform as an internal product, its
paved road, its SLOs and its adoption. Here, the decision to create the team and what load is
transferred to it.
itsm-itil-standards: service, catalogue and continuous operation.
grc-compliance-standards: regulatory framework, formal evidence, segregation of duties.
ai-agent-workflow-standards (already written): the team policy on coding
agents is theirs — which task is delegated, instruction files, permissions, review of the
diff, attribution. Here only the decision to adopt it, who pays its organisational cost and which
metrics are NOT used to justify it (§3.11).
knowledge-management-standards: where documentation lives and how it is
kept alive. Here, the obligation that the decision be written down.
2. Default decisions
Verify on the web the status of the cited sources before relying on them (§8).
| Decision |
Default |
Justifiable alternative |
| What distinguishes one level from another |
Scope of impact, type of decision and time horizon (§3.1) |
Never years of experience and never mastery of a technology |
| Career track |
Two tracks with parity of level, scope and pay, published |
A single track only in organisations with < 15 engineers, and declaring it |
| Staff+ work assignment |
By declared archetype (§3.2), agreed with the person and revisable |
— |
| Record of technical decisions |
ADR mandatory, Nygard format, in the repository of the affected code |
A central record only if the decision crosses several repositories |
| Decision speed |
Derived from reversibility (§3.3): reversible → fast and at the lowest level; irreversible → informed and higher up |
— |
| Before building anything non-trivial |
A reviewed design document with acceptance criteria (§3.4) |
A throwaway prototype with a time box, if the problem is that it is not understood |
| Technical debt |
An explicit and recurring budget line, with its unit of measure (§3.5) |
Never "when there is time" |
| Measure of a technical leader |
The team's result: delivery, reliability, autonomy, recorded decisions |
— |
| The leader's contact with the code |
The minimum sufficient to judge (§3.8): reviewing, reading the design, touching the system periodically |
— |
| Individual productivity metric |
It does not exist. Forbidden (§7) |
Qualitative evaluation with written and dated evidence |
| 1:1 cadence |
Weekly or fortnightly, the report's agenda, never a status report (§3.9) |
— |
| Team size |
Stable, with bounded load and a single owner for each system (§3.6) |
— |
| Technical disagreement |
Resolved with data, with a deadline, and closed with disagree-and-commit (§3.7) |
Escalation only after exhausting the bounded experiment |
Sources: what is cited and with what caveat
- ✅ One-way / two-way doors. Primary source: Jeff Bezos, 2015 letter to Amazon
shareholders (
ir.aboutamazon.com, PDF). Verbatim (extracted from the official PDF,
not from a summary): "Some decisions are consequential and irreversible or nearly
irreversible – one-way doors – and these decisions must be made methodically, carefully, slowly,
with great deliberation and consultation. If you walk through and don't like what you see on the
other side, you can't get back to where you were before. We can call these Type 1 decisions. But
most decisions aren't like that – they are changeable, reversible – they're two-way doors. If
you've made a suboptimal Type 2 decision, you don't have to live with the consequences for that
long. You can reopen the door and go back through. Type 2 decisions can and should be made quickly
by high judgment individuals or small groups." And the diagnosis that matters here: "As
organizations get larger, there seems to be a tendency to use the heavy-weight Type 1
decision-making process on most decisions, including many Type 2 decisions. The end result of this
is slowness, unthoughtful risk aversion, failure to experiment sufficiently, and consequently
diminished invention." Declared discrepancy: numerous secondary sources place the framework
in the 2016 letter; that is incorrect — the Type 1 / Type 2 terminology appears in the 2015 one.
The 2016 one contains related material (disagree and commit), which is a different thing.
- ✅ ADR. Primary source: Michael Nygard, "Documenting Architecture Decisions", 15-Nov-2011,
originally published on the Relevance/ThinkRelevance blog and today at
cognitect.com. A five-section
template: Title (a short noun phrase with a sequential number that is never reused),
Status (proposed / accepted / deprecated / superseded by ADR-NNNN), Context (the forces
at play, in neutral language), Decision, Consequences (positive and negative). Derived
rule: an accepted ADR is not edited; if the conclusion changes, a new one is written that
supersedes the previous one and updates its status.
- ⚠️ Staff+ archetypes. A published reference taxonomy does exist: Will Larson, Staff
Engineer: Leadership beyond the management track (2021) and its origin in the open
(
lethain.com/staff-engineer-archetypes/, staffeng.com/guides/staff-archetypes/). Four
archetypes: tech lead, architect, solver, right hand. Explicit caveat: it is
a descriptive taxonomy observed in fast-growing technology companies, not a
research result, and it has published criticism from practitioners (Sean Goedecke: the
archetypes exist but are bad advice as a career target, because solver and right hand
depend on accumulated trust and cannot be pursued directly; Alex Ewerlöf: they should not
be used as job titles). Permitted use: vocabulary for distributing work and clarifying
expectations. Forbidden use: turning them into levels, titles or boxes on a ladder.
- ⚠️ Team cognitive load. Original source John Sweller, "Cognitive load during problem
solving: Effects on learning", Cognitive Science 12(2):257-285 (1988) — a theory of
individual learning and instructional design. Its transfer to teams is from Matthew Skelton and
Manuel Pais, Team Topologies (2019). Treat it as a useful analogy, not as a validated measurement,
exactly as in
platform-engineering-standards §6.4: team cognitive load has
no measurable unit, the sum of individual loads is not a construct defined in the original
theory, and DevOps academic literature points out that the book provides no scientific evidence
and rests on experience and cases. It is useful for deciding how systems are distributed between teams; it is
not useful as a number on a slide.
- ✅ SPACE. Nicole Forsgren, Margaret-Anne Storey, Chandra Maddila, Thomas Zimmermann, Brian
Houck and Jenna Butler, "The SPACE of Developer Productivity: There's more to it than you think",
ACM Queue 19(1):20-48 (2021), DOI 10.1145/3454122.3454124 (also in CACM). Five
dimensions: satisfaction and well-being, performance, activity, communication and collaboration,
efficiency and flow. The operational rule taken from it: measure at least three dimensions at
once, combining objective metrics and a survey; activity never in isolation to reward
or penalise. Declared discrepancy and usage warning: the claim circulates that "SPACE
measures the individual and DORA the team" — that is a misreading: one of the myths the article
itself dismantles is that productivity is only individual performance, and its central thesis is
that it is not captured by a single metric nor by activity data. It has not been possible to retrieve the
full text verbatim (ACM Queue and CACM return 403 to automated download): the
formulations above come from consistent secondary sources and must be checked against
the original article before quoting them literally (§8).
- ✅ DORA as a system measure. The four metrics measure team and system delivery,
not people. Report 2025: State of AI-assisted Software Development (
dora.dev/dora-report-2025/,
~5,000 professionals); report 2026: The ROI of AI-Assisted Software Development (InfoQ
coverage, May-2026). Caveat about the 2026 one: there is published criticism that, compared with the
research rigour of previous editions, it is largely a set of speculative
recommendations. Cite the edition and the year; never a bare "DORA says".
Famous figures: what is NOT used as data
This domain is saturated with numbers that circulate with no locatable primary study or with
dismantled methodology. These are expressly discarded:
- ❌ The "10x programmer". Origin: Sackman, Erikson and Grant, "Exploratory Experimental Studies
Comparing Online and Offline Programming Performance", CACM 11(1):3-11, Jan-1968. Methodological
refutation: Lutz Prechelt, "The 28:1 Grant/Sackman legend is misleading, or: How large is
interpersonal variation really?", Technical Report 1999-18, Universität Karlsruhe, Dec-1999
(
page.mi.fu-berlin.de/prechelt/Biblio/varianceTR.pdf). Flaws in the original: 12 subjects,
trivial tasks of a mathematical nature, and a mix of low-level and high-level language
programmers in the same group, which inflates the best/worst ratio. Prechelt reanalyses a much
larger dataset and argues that comparing the best with the worst is the wrong comparison: against
the median of the worst quartile versus that of the best, the ratio rarely exceeds 4, and the
standard deviation/mean ratio is around 0.5. Permitted use: "there is substantial
interpersonal variation, of the order of 2-4x depending on the task, with overlapping distributions". Forbidden use:
"10x", and above all using it to justify pay, staffing or redundancies.
- ❌ "An interruption costs 23 minutes and 15 seconds". The paper that backs it does not
exist. The paper that gets cited, Mark, González and Harris, "No Task Left Behind? Examining the
Nature of Fragmented Work", CHI '05, measures the probability of resuming a task on the same day
and does not contain that figure. The other one usually cited, "The Cost of Interrupted Work: More
Speed and Stress", finds that interrupted subjects completed the tasks faster,
with more stress. The figure comes from a 2006 interview with Gloria Mark, not from a peer-reviewed
article. It is not written, not even with caveats. What is defensible without a figure: fragmentation has a cost
and the argument for protecting blocks of work does not need an invented number.
- ❌ "70 % of the variance in team engagement is determined by the manager" (Gallup, State
of the American Manager, 2015). Proprietary analysis, not peer reviewed, with no raw data
published, and with two interpretation problems that its diffusion ignores: it refers to variance
between teams (it is not an individual-level claim), and the "manager effect" absorbs everything
that clusters at team level (function, location, load, staffing). It is not used as data to
justify a decision.
- ❌ "People leave managers, not companies" (First, Break All the Rules, Gallup). There is a published
refutation from Culture Amp (2017) over 175 teams: people do leave bad managers, but it is not
the first reason, and in bad organisations having a good or bad manager barely changes the decision to
leave. The weak version holds ("the manager's quality is one of the factors, conditional on
the organisation being decent"); the slogan does not.
- ❌ Turnover percentages attributed to a "bad manager" and the cost of a bad hire
expressed as a multiple of salary (§
technical-hiring-standards): discarded for the same
reason, there in detail.
General rule: a figure with no primary study, year, sample and method does not go into a document,
a slide or a budget request. If the argument only stands with the figure, the
argument was bad.
3. Structure and conventions
3.1 Career tracks: what actually changes
What changes when you move up a level is not "the level": it is three observable variables.
| Variable |
Senior |
Staff |
Principal |
| Scope of impact |
Their team and the system they answer for |
Several teams, or a critical area the organisation depends on |
The engineering organisation or a cross-cutting capability |
| Type of decision |
How it is implemented; chooses between known options |
Which options exist; defines the problem and the interfaces |
What is stopped; irreversible decisions and trade-offs between areas |
| Time horizon |
Weeks to a quarter |
Quarters to a year |
One to several years |
| Required evidence |
Systems delivered and operable |
Other people's designs improved, decisions recorded, people levelled up |
Changes of direction the organisation followed and that can be traced |
Hard rules:
- The ladder is published. A ladder only the manager knows is not a ladder, it is a retroactive
excuse. Public references exist so as not to start from zero — Rent the Runway (2015, with
Camille Fournier as CTO), Dropbox, CircleCI, Square, Kickstarter, aggregated at
progression.fyi and github.com/bmoeskau/engineering-ladders. They are adapted, not copied: someone
else's ladder describes someone else's organisation.
- Real parity between tracks. The individual contributor track and the management one have the same
level ceiling, the same salary band and the same presence in decision-making forums. Without those
three, the technical track is decorative and everyone knows it within six months.
- Managing is not a promotion, it is a change of job. The work, the measure and the skill are
different. Operational corollary: there is a declared way back, and using it is not a failure
recordable in the evaluation.
- The level is assigned on written evidence, not on seniority and not on likeability. If the evidence cannot
be written down with dated facts, there is no level.
- No level is defined by a technology. "Kubernetes senior" is not a level, it is a
dependency on a tool with an expiry date.
3.2 Staff+ archetypes: what they are really for
They serve one purpose, and that is why they are here: avoiding assigning work the person cannot
do from where they are. Taxonomy and caveats in §2.
| Archetype |
Work that belongs to them |
Work that destroys their impact |
| Tech lead |
Approach and execution of one team, alongside the manager |
Being the bottleneck for the decisions of another three teams |
| Architect |
Direction, quality and approach of one critical area, with a multi-year horizon |
Firefighting outside their area; making their area depend on their presence |
| Solver |
Going into a hard, bounded problem and coming out with the path resolved |
Becoming the permanent owner of what they fixed |
| Right hand |
Extending an executive's reach in a large organisation |
Existing in an organisation that does not need one: it reproduces hierarchy without adding judgement |
Rules:
- The archetype is agreed in writing with the person and is reviewed (by default, every six months or
when the project changes). A solver who gets handed the maintenance of everything they touched
stops being a solver within a quarter.
- Architect and right hand only appear at a certain scale. Larson observes them emerging around
~100 and ~1,000 engineers respectively. Creating them earlier manufactures a decision layer with no
problem to solve.
- FORBIDDEN to use them as titles or as levels on the ladder (§7).
3.3 Decisions: what is decided, who decides and how fast
Who decides, by default:
| Type of decision |
Decided by |
Consulted |
Recorded in |
| Implementation inside a module of the team's own |
Whoever implements it |
The PR reviewer |
The code and the PR |
| Choice of library or pattern within the team |
The team, the tech lead arbitrates |
— |
The repository's ADR |
| Contract or interface between two teams |
Both teams, jointly |
The area's architect |
ADR + versioned contract (api-design-standards) |
| A new technology in the organisation, or the retirement of an existing one |
Staff+ level or architecture, with published criteria |
Platform, security, FinOps |
ADR + inventory |
| A scope or date commitment |
Not this skill |
— |
project-management-standards |
| Acceptance of a regulatory risk |
Whoever has the formal authority |
— |
grc-compliance-standards |
The ADR is an obligation, not a courtesy. Minimum rule: if a technical decision is going to condition the
work of someone who was not in the conversation, it gets written. Nygard format (§2), in the
affected repository, numbered and not editable once accepted.
Speed is set by reversibility, not by perceived importance (Bezos 2015, verbatim in §2):
|
One-way door (Type 1) |
Two-way door (Type 2) |
| Typical examples |
Public data model, identity scheme, API format exposed to third parties, lossy migration, choice of a vendor with an expensive exit |
Internal library, folder structure, internal log format, test framework |
| Speed |
Slow, deliberate, informed: design document, alternatives evaluated, formal review |
Fast, by the person or the small group closest to the problem |
| Level that decides |
Higher up, with consultation |
The lowest possible |
| Cost of being wrong |
High and permanent |
Low: the door reopens |
- Every ADR carries a
reversibilidad field with an explicit value. Without it, the organisation treats
everything as Type 1 and grinds to a halt — the failure the 2015 letter itself describes.
- The classification can be challenged, but in writing and with an argument. "This is irreversible"
is a falsifiable claim: you ask for the estimated cost of reverting it.
- A Type 2 decision that has been under discussion for three weeks already cost more than being wrong. Rule:
a time box, and if it expires, the person with the responsibility decides and it is recorded.
3.4 Design document and design review
The design review is the point in the cycle where a mistake is still cheap. Afterwards there is
code, dependencies, migrated data and people who have already defended it in public.
When a design document is mandatory (any one is enough):
- The decision crosses a team boundary or creates/changes an interface between teams.
- It is a one-way door (§3.3).
- It touches personal data, authentication, authorisation, cryptography or money.
- The estimated effort exceeds a threshold the team publishes (by default: two person-weeks).
- A technology the organisation does not yet operate is introduced.
Minimum template. A document without sections 5, 6 and 8 is sent back unreviewed.
# Diseño: <nombre> Autor: <persona> Estado: borrador|en revisión|aceptado|sustituido
1. Problema # qué falla hoy, con evidencia observable (dato, incidente, ticket)
2. Restricciones # plazo, presupuesto, normativa, compatibilidad, equipo disponible
3. No objetivos # lo que este diseño explícitamente NO resuelve
4. Propuesta # la solución, con el diagrama mínimo que la explique
5. Alternativas descartadas # >=2, con el motivo del descarte. Sin esto no hay diseño, hay preferencia
6. Criterios de aceptación # falsables y medibles: cómo se sabrá que funcionó, y cuándo se mira
7. Impacto # migración, compatibilidad hacia atrás, coste recurrente, operabilidad, seguridad
8. Reversibilidad # Tipo 1 o Tipo 2, y coste estimado de revertir
9. Riesgos abiertos # con dueño
Review rules:
- The problem is reviewed before the solution. If the reviewers do not agree on the
problem, discussing the solution is wasted time and it always ends in aesthetics.
- Mandatory pre-reading, with a deadline (48 h by default). A meeting where the document is read
live is a reading meeting, not a review.
- Comments are labelled as blocking or non-blocking, just as in code review
(
code-review-standards). A reviewer who marks nothing as blocking has approved.
- The review ends with a decision and a date, not with "let's keep talking". The
valid states are: accepted, accepted with listed conditions, rejected with a reason, or postponed with
a date and with what remains to be found out.
- The accepted document becomes an ADR or links to one. A design that leaves no trace of a decision
gets discussed again in six months.
- Acceptance criteria with a check date. They are revisited when due: if they were not met,
that is information about how the team designs, and it is the cheapest input there is for improving.
3.5 Technical debt: a business decision with a budget
The technical criteria —what to refactor, with which technique and with which test net— belong to
refactoring-tech-debt-standards (already written). Here only the decision to invest and
how it is defended.
- Debt is recorded like any other work, in the same backlog, with an owner and with the
cost it imposes today — not with adjectives. Admissible formulation: "every change in the
billing module requires touching four places and last quarter it generated 3 of the 7 incidents".
Inadmissible formulation: "the code is bad".
- Distinguish deliberate debt from degradation. Deliberate debt was incurred with a recorded
decision and a review date (if it does not have one, it was not a decision: it was an oversight).
Degradation is silent accumulation and is detected by symptom, not by opinion.
- An explicit and recurring budget. A percentage of capacity per cycle is declared,
published and protected the way a committed date is protected. A percentage that gets cancelled in
the first quarter under pressure was not a budget: it was an intention.
- How it is defended to whoever pays — and this is what always fails. Technical debt is not
defended as code quality: it is defended with the cost the business is already paying, in
its vocabulary. Arguments that work because they are verifiable:
- Lead time: the same change takes X in this module and Y in the rest.
- Reliability: the proportion of incidents concentrated in the component
(
incident-management-standards provides the datum).
- Direct cost: oversized infrastructure, licences, hours of manual operation
(
finops-standards).
- Risk with a name: an unsupported dependency, a version with no security patches
(
vulnerability-management-standards), regulatory non-compliance (grc-compliance-standards).
- The option that is lost: what will not be doable, and when, if it is not touched.
- A refactor request with none of those five arguments is rejected, whoever makes
it. And it is rejected here, inside engineering, before the business rejects it: credibility
is spent once.
- Never "a refactor release" that delivers nothing. Debt work is delivered in
increments with an observable effect; if it cannot be split, the problem is the design of the
intervention, not the calendar.
3.6 The team as a system
- A single owner per system. A system with no named owner is maintained by whoever had the bad luck
of touching it last. Two owners is zero owners.
- Bounded load, in the analogical sense of cognitive load (§2, with its caveat): a team
answers for as many systems as it can understand, operate and improve. An operational and
falsifiable test that requires no invented metric: can the team deploy, diagnose in
production and explain the data model of every system it answers for, without depending on one
specific person? If not, the load exceeds the team — the scope is reduced or extraneous load
is transferred to the platform (
platform-engineering-standards).
- Stable teams. Reorganising teams resets system knowledge and trust
relationships. Every reorganisation is justified in writing with the problem it solves and its expected
cost, or it is not done.
- Size: small enough that everyone knows everyone's work, large enough
to sustain an on-call rota without burning anyone (
sre-practice-standards fixes the latter).
Adding people to a saturated team adds coordination before it adds capacity
(project-management-standards §6).
- Inter-team dependencies: they are managed by interface, not by meeting. Preference, in order:
(1) eliminate the dependency, (2) decouple with an interface contract and a simulator, (3) sequence
with a mutual commitment from the providing team, (4) escalate. Adding a recurring
sync meeting is choosing (4) and calling it (3).
- The team's structure and the system's converge (Conway). Practical consequence: if you
want a different architecture, you have to change the team boundaries, not just the
diagram.
3.7 Technical conflict
A technical disagreement with no procedure gets resolved by hierarchy or by exhaustion, and both
produce the worst decision available. Procedure, in this order and with a deadline:
- Make the disagreement explicit: each side writes, in one paragraph, what they defend and what evidence
would make them change their mind. Whoever cannot answer the second has no technical position,
they have a preference — and preferences do not block.
- Find the datum: prototype, measurement, load test, review of a past incident. With
a fixed time box agreed beforehand.
- If the datum does not discriminate (a real tie), whoever answers for the consequences decides —
normally the owner of the affected system — and it is recorded in an ADR with the alternatives and the
tie declared.
- Disagree and commit, as an explicit team rule. Reference: Amazon's leadership
principle "Have Backbone; Disagree and Commit" (
amazon.jobs) — the obligation to challenge the
decision respectfully even when it is uncomfortable, and full commitment once it is taken. The term is
older than Amazon (Intel, constructive confrontation). The local rule, which is what applies:
once the decision is closed, nobody passively sabotages it or reopens it in corridors; it is reopened
only with new evidence and through the same channel.
- When it is escalated: when the disagreement crosses the boundary of two teams and blocks work for more
than a week, or when it involves a security, legal or personal-data risk. Escalating is not
losing: escalating late is.
- A disagreement that repeats is an organisational design problem, not a people problem: the
boundaries of responsibility are badly drawn (§3.6).
3.8 The technical leader's work that is not writing code
Their value is measured in the team's result, not in their own output. Direct consequences:
if the most productive person on the team is the leader, the team is underused; if the leader
is on the critical path of every delivery, they are a SPOF with holidays.
Work that is theirs, and that produces leverage:
- Unblocking: identifying who is stuck and why, and getting it out of the way. It is the highest-return
activity of the job and the first one sacrificed when the leader starts programming.
- Defining interfaces between teams before two teams build against different
assumptions (§3.6).
- Reviewing other people's designs and decisions (§3.4) — not rewriting them.
- Writing what the organisation will need to remember: ADRs, context, criteria.
- Creating the conditions for someone else to take the decision, and accepting that they take it differently from how
the leader would have taken it if the result is acceptable. A leader who only delegates the decisions that
would have matched their own has delegated nothing.
- Developing people (§3.9): it is work with a calendar, not a by-product.
The opposite trap, and it is just as real: the leader who stops touching the code loses the ability
to judge. With no contact with the system, other people's estimates become uncheckable, designs
get approved out of trust in the person and not because of the content, and technical debt becomes
a story with no evidence. Non-negotiable operational minimum:
- Review code properly and regularly (
code-review-standards), including changes they do not
fully understand, which is where you learn where the system is.
- Be in the on-call rotation if the team has one, or at least take part in incident
response (
incident-management-standards).
- Take real work, but never on the critical path: internal tooling, fixes,
tests, the task nobody wants. FORBIDDEN to take the quarter's critical feature:
it turns them into a bottleneck and leaves the team without the developmental part.
- Arbitration rule between the two traps: if the team delivers just as well when the leader is
away for two weeks, the split is correct. It is the only test that matters, and it can be run
literally.
Code review is not used as an instrument of power. Vetoed manifestations: blocking a
PR out of aesthetic preference, demanding your own solution with no technical argument, withholding approval
as leverage in a negotiation unrelated to the diff. The criteria for what blocks belong to
code-review-standards; the obligation that the leader does not distort it belongs here.
3.9 1:1s, feedback and evaluation
The 1:1 belongs to the report. Concrete rules:
- The agenda is set by whoever reports, and it exists before the meeting. If there is no agenda, the default
question is not "how is X going?" but "what is getting in your way most?".
- FORBIDDEN for it to be a status report: the state of the work is in the management system
(
project-management-standards) and consulting it is the leader's responsibility, not minutes of the
person's time.
- A fixed and protected cadence (weekly or fortnightly by default). Cancelling it repeatedly
communicates a priority more clearly than any speech.
- Shared notes and commitments with an owner and a date. A 1:1 with no trace repeats identically for
a year.
- At least once a quarter, the 1:1 is about career: current level, missing evidence,
archetype (§3.2), what work needs to be sought to get there. With the published ladder in front of you.
Feedback:
- Specific, dated and about observable behaviour, not about traits. "In the review of the payments
design on 12 May you closed the discussion before Ana presented her alternative" is
actionable. "You are not very collaborative" is not: what cannot be observed cannot be corrected.
- Close to the event. Feedback saved for six months for the annual review is not feedback: it is an
ambush, and it destroys the credibility of the whole process.
- No surprises in the evaluation. Falsifiable rule: if something appears in the annual review and it is
the first time the person hears it, the failure is the leader's, and it is recorded as such.
- Difficult feedback is given in private, in writing as well as verbally, and with what is expected
differently and by when.
Why individual engineering productivity metrics are counterproductive — and this
is the part you have to be able to defend to management with the source in hand:
- The published position of the sector's reference research. DORA: the four metrics
measure team and system delivery and must not be used to evaluate the performance of
individual engineers; applying them to people creates perverse incentives. SPACE (Forsgren et
al., 2021, §2) dismantles two directly applicable myths: that productivity is activity, and
that it is only individual performance; and it establishes that activity metrics are never used
in isolation to reward or penalise.
- They are trivial to game and gaming them is rational: lines, commits, PRs and points are
free to inflate and the cost of inflating them is paid by someone else (the reviewer, the maintainer).
- They measure the visible and punish the valuable: reviewing, pairing, mentoring, being on call,
deleting code and preventing the unnecessary from being built leave no trace in any counter.
- Engineering work is interdependent: attributing the result to an individual within
a system with queues, dependencies and review is an attribution error, not an imprecise
measurement.
- The alternative that is used: qualitative evaluation with written and dated evidence against
the published ladder (§3.1) — designs, recorded decisions, incidents resolved, people
developed, systems that work without their author. It is more work for the leader. That is the
cost of the job.
3.10 Incidents: what corresponds to leadership
The full process belongs to incident-management-standards. Here only the part that only
leadership can do, and which is where the blameless culture breaks in practice:
- Protecting the blameless postmortem when there is pressure from above to name someone. The
pressure is real and arrives in the form of a reasonable question ("who deployed it?"). The standard
answer, in the same language as whoever asks: the system allowed an individual error
to reach production; that is the defect and that is what gets fixed, because it is the only one that will not
recur. Replacing the person leaves the system identical for the next one.
- Verifiable drafting rule: the postmortem contains no proper names as a cause. If
"X ran Y" appears, it is rewritten as "the procedure allowed Y to be run without confirmation".
- Honest contraindication: blameless does not mean without responsibility. Repeated
negligence, knowingly bypassing controls or acting without authorisation are a conduct matter and are
handled outside the postmortem, through another channel and with HR. Mixing them destroys both.
- The postmortem's actions get funded. A postmortem whose actions do not enter the next
cycle's plan teaches the team that the exercise is theatre, and from then on postmortems
are written to be filed.
- The leader publicly owns the team's failure and publicly attributes the success. It is a
conduct rule with a verifiable effect: without it, nobody reports a problem early.
3.11 Coding agents: what leadership decides
The concrete policy —which task is delegated, instruction files, permissions, review of the diff,
attribution— belongs to ai-agent-workflow-standards. **Here, three
…(truncated)
1---2name: tech-leadership-standards3description: Technical leadership as a set of decisions and artifacts, not a personality trait. Use when defining or applying an engineering career ladder and the IC-versus-manager track, assigning work by Staff+ archetype (tech lead, architect, solver, right hand), writing or reviewing an ADR and classifying a decision as a one-way or two-way door, writing a design document and running a design review with acceptance criteria, deciding who decides and at what level, budgeting technical debt as an explicit business decision, protecting a blameless postmortem under management pressure, running a 1:1 with a report-owned agenda, giving written performance feedback, being asked for individual productivity metrics such as lines of code, commits, velocity or DORA per engineer, citing SPACE or DORA in a measurement argument, sizing a team and its cognitive load in the Team Topologies sense, mapping inter-team dependencies, resolving a technical disagreement and applying disagree-and-commit, deciding whether to adopt coding agents and who pays the organizational cost, or being handed responsibility without authority.4metadata:5 source: https://skillmd.com/skills/serialexperimentslainnnn/tech-leadership-standards6---78# Technical leadership standards910Criteria verified as of **August 2026**. Re-verify on the web before committing to anything (§8).1112## 1. Scope and triggers1314Covers **the decisions taken by whoever leads technically and the artifacts in which they are15written down**: career tracks and the real scope of each level, staff+ role archetypes, who decides what16and how fast, the decision record, the design document and the design review, technical debt17as a budget line, the individual conversation and performance evaluation, team18sizing and its dependencies, and the resolution of technical disagreements.1920Triggers: "career ladder", "level ladder", "senior/staff/principal", "IC vs. management",21"staff archetype", "tech lead", "ADR", "decision record", "one-way door",22"reversible", "design document", "design doc", "design review", "technical RFC", "technical23debt", "refactor budget", "blameless postmortem", "1:1", "feedback", "performance24review", "individual productivity metrics", "lines of code", "commits per person",25"DORA per engineer", "SPACE", "cognitive load", "Team Topologies", "team size",26"inter-team dependencies", "technical disagreement", "disagree and commit", "responsibility without27authority".2829**Guiding principle**: **technical leadership is a set of decisions and artifacts, not a character30trait.** It is not evaluated by how the person is perceived, but by what exists in writing and by31what the team achieves: decisions recorded with their reversibility, designs reviewed before32building, interfaces between teams defined, debt with an assigned budget and people with33specific, dated feedback. **A falsifiable test applicable to any claim about leadership:34name the artifact it produces, who reads it and which decision it changes.** If it cannot be named, the35claim is motivational and does not get written here.3637Hard corollary: **"lead by example", "foster trust" and phrases of the same genre fix38nothing.** They are not criteria, they are decoration. Their actionable equivalent always exists and is the only thing39this document accepts: *"every architecture decision lives in an ADR with declared40reversibility"*, *"the postmortem names no people and the leader defends it to management in writing"*.4142**Not applicable**:43- `technical-hiring-standards`: **reciprocal and strict**. Here you decide **what profile is needed,44 why, and which gap in the team it fills** (§3.7); there **how a candidate is measured** — rubric,45 exercise format, method validity, bias, legal regime and onboarding. One-sentence boundary:46 **leadership defines the role; the selection process is the measuring instrument.**47- `project-management-standards` (**already written**): **delivery, the plan, the commitment, the48 estimate, the risk and the stakeholders are theirs**, including the record of *management*49 decisions. Here the **technical decision** and **team development**. Boundary: if the question50 is *"when will it be ready and with what risk?"*, it is theirs; if it is *"how is it built and who decides it?"*,51 it is ours. **The decision record format is the same in both** (§3.3) so as not to fork52 the repository.53- `product-discovery-standards`: **what gets built and why** — problem,54 user, hypothesis, prioritisation by value and the decision to kill an idea. Here the what is not decided.55- `enterprise-architecture-standards`: architecture governance for the56 organisation, capability roadmap and application portfolio. Here the scope is the team57 and its immediate neighbourhood, not the company. Reciprocal declared in their §1.58- `software-architecture-patterns-standards` and `refactoring-tech-debt-standards` (**already written**):59 **the technical criteria are theirs** — which pattern applies, when a module is refactored60 and with which technique. Here only **the decision to invest in it, its budget and how it is defended**61 (§3.5). Both already declare the reciprocal in their §1.62- `microservices-architecture-standards`: service boundaries and concrete contracts. Here, the63 organisational reflection of those boundaries (§3.6).64- `code-review-standards` (**already written**): **the review criteria are theirs** — what blocks, how65 the comment is written, what a high-risk change requires. Here only the rule that **review66 is not used as an instrument of power** (§3.8) and who arbitrates the disagreement that review67 does not close.68- `testing-qa-standards`: test strategy and gates. Here, the decision to fund it.69- `sre-practice-standards`: **SLOs, *error budget*, on-call, capacity and the DORA metrics are70 theirs**. Here, solely, that **DORA is a system measure and its individual use is forbidden**71 (§3.9, §7).72- `incident-management-standards`: **the incident process and the blameless postmortem are theirs**73 — roles, timeline, actions. Here what corresponds to leadership and only that: **protecting the74 blameless postmortem when there is pressure from above to name a culprit** (§3.10).75- `platform-engineering-standards` (**already written**): **the platform as an internal product**, its76 paved road, its SLOs and its adoption. Here, the decision to create the team and what load is77 transferred to it.78- `itsm-itil-standards`: service, catalogue and continuous operation.79- `grc-compliance-standards`: regulatory framework, formal evidence, segregation of duties.80- `ai-agent-workflow-standards` (**already written**): **the team policy on coding81 agents is theirs** — which task is delegated, instruction files, permissions, review of the82 diff, attribution. Here only **the decision to adopt it, who pays its organisational cost and which83 metrics are NOT used to justify it** (§3.11).84- `knowledge-management-standards`: where documentation lives and how it is85 kept alive. Here, the obligation that the decision be written down.8687## 2. Default decisions8889> Verify on the web the status of the cited sources before relying on them (§8).9091| Decision | Default | Justifiable alternative |92|---|---|---|93| What distinguishes one level from another | **Scope of impact, type of decision and time horizon** (§3.1) | Never years of experience and never mastery of a technology |94| Career track | **Two tracks with parity of level, scope and pay**, published | A single track only in organisations with < 15 engineers, and declaring it |95| Staff+ work assignment | **By declared archetype** (§3.2), agreed with the person and revisable | — |96| Record of technical decisions | **ADR mandatory**, Nygard format, in the repository of the affected code | A central record **only** if the decision crosses several repositories |97| Decision speed | **Derived from reversibility** (§3.3): reversible → fast and at the lowest level; irreversible → informed and higher up | — |98| Before building anything non-trivial | **A reviewed design document** with acceptance criteria (§3.4) | A throwaway prototype with a time box, if the problem is that it is not understood |99| Technical debt | **An explicit and recurring budget line**, with its unit of measure (§3.5) | Never "when there is time" |100| Measure of a technical leader | **The team's result**: delivery, reliability, autonomy, recorded decisions | — |101| The leader's contact with the code | **The minimum sufficient to judge** (§3.8): reviewing, reading the design, touching the system periodically | — |102| Individual productivity metric | **It does not exist.** Forbidden (§7) | Qualitative evaluation with written and dated evidence |103| 1:1 cadence | **Weekly or fortnightly, the report's agenda**, never a status report (§3.9) | — |104| Team size | **Stable, with bounded load and a single owner for each system** (§3.6) | — |105| Technical disagreement | **Resolved with data, with a deadline, and closed with disagree-and-commit** (§3.7) | Escalation **only** after exhausting the bounded experiment |106107### Sources: what is cited and with what caveat108109- ✅ **One-way / two-way doors.** Primary source: **Jeff Bezos, 2015 letter to Amazon110 shareholders** (`ir.aboutamazon.com`, PDF). **Verbatim** (extracted from the official PDF,111 not from a summary): *"Some decisions are consequential and irreversible or nearly112 irreversible – one-way doors – and these decisions must be made methodically, carefully, slowly,113 with great deliberation and consultation. If you walk through and don't like what you see on the114 other side, you can't get back to where you were before. We can call these Type 1 decisions. But115 most decisions aren't like that – they are changeable, reversible – they're two-way doors. If116 you've made a suboptimal Type 2 decision, you don't have to live with the consequences for that117 long. You can reopen the door and go back through. Type 2 decisions can and should be made quickly118 by high judgment individuals or small groups."* And the diagnosis that matters here: *"As119 organizations get larger, there seems to be a tendency to use the heavy-weight Type 1120 decision-making process on most decisions, including many Type 2 decisions. The end result of this121 is slowness, unthoughtful risk aversion, failure to experiment sufficiently, and consequently122 diminished invention."* **Declared discrepancy**: numerous secondary sources place the framework123 in the **2016** letter; that is incorrect — the Type 1 / Type 2 terminology appears in the **2015** one.124 The 2016 one contains related material (*disagree and commit*), which is a different thing.125- ✅ **ADR.** Primary source: **Michael Nygard, "Documenting Architecture Decisions", 15-Nov-2011**,126 originally published on the Relevance/ThinkRelevance blog and today at `cognitect.com`. A five-section127 template: **Title** (a short noun phrase with a sequential number that is never reused),128 **Status** (proposed / accepted / deprecated / superseded by ADR-NNNN), **Context** (the forces129 at play, in neutral language), **Decision**, **Consequences** (positive and negative). Derived130 rule: **an accepted ADR is not edited**; if the conclusion changes, a new one is written that131 supersedes the previous one and updates its status.132- ⚠️ **Staff+ archetypes.** A published reference taxonomy does exist: **Will Larson, *Staff133 Engineer: Leadership beyond the management track* (2021)** and its origin in the open134 (`lethain.com/staff-engineer-archetypes/`, `staffeng.com/guides/staff-archetypes/`). Four135 archetypes: **tech lead**, **architect**, **solver**, **right hand**. **Explicit caveat**: it is136 a **descriptive taxonomy observed in fast-growing technology companies, not a137 research result**, and it has published criticism from practitioners (Sean Goedecke: the138 archetypes exist but are bad advice as a career target, because *solver* and *right hand*139 depend on accumulated trust and cannot be pursued directly; Alex Ewerlöf: they should not140 be used as job titles). **Permitted use: vocabulary for distributing work and clarifying141 expectations. Forbidden use: turning them into levels, titles or boxes on a ladder.**142- ⚠️ **Team cognitive load.** Original source **John Sweller, "Cognitive load during problem143 solving: Effects on learning", *Cognitive Science* 12(2):257-285 (1988)** — a theory of144 **individual learning and instructional design**. Its transfer to teams is from **Matthew Skelton and145 Manuel Pais, *Team Topologies* (2019)**. **Treat it as a useful analogy, not as a validated measurement**,146 exactly as in `platform-engineering-standards` §6.4: team cognitive load **has147 no measurable unit**, the sum of individual loads is not a construct defined in the original148 theory, and DevOps academic literature points out that the book **provides no scientific evidence**149 and rests on experience and cases. It is useful for **deciding how systems are distributed between teams**; it is150 not useful as a number on a slide.151- ✅ **SPACE.** **Nicole Forsgren, Margaret-Anne Storey, Chandra Maddila, Thomas Zimmermann, Brian152 Houck and Jenna Butler, "The SPACE of Developer Productivity: There's more to it than you think",153 *ACM Queue* 19(1):20-48 (2021), DOI 10.1145/3454122.3454124** (also in *CACM*). Five154 dimensions: satisfaction and well-being, performance, activity, communication and collaboration,155 efficiency and flow. The operational rule taken from it: **measure at least three dimensions at156 once, combining objective metrics and a survey**; activity **never in isolation** to reward157 or penalise. **Declared discrepancy and usage warning**: the claim circulates that "SPACE158 measures the individual and DORA the team" — **that is a misreading**: one of the myths the article159 itself dismantles is that productivity is *only* individual performance, and its central thesis is160 that it is not captured by a single metric nor by activity data. **It has not been possible to retrieve the161 full text verbatim** (ACM Queue and CACM return 403 to automated download): the162 formulations above come from consistent secondary sources and **must be checked against163 the original article before quoting them literally** (§8).164- ✅ **DORA as a system measure.** The four metrics measure **team and system delivery**,165 not people. Report **2025: *State of AI-assisted Software Development*** (`dora.dev/dora-report-2025/`,166 ~5,000 professionals); report **2026: *The ROI of AI-Assisted Software Development*** (InfoQ167 coverage, May-2026). **Caveat about the 2026 one**: there is published criticism that, compared with the168 research rigour of previous editions, it is largely a set of speculative169 recommendations. Cite the edition and the year; **never a bare "DORA says"**.170171### Famous figures: what is NOT used as data172173This domain is saturated with numbers that circulate with no locatable primary study or with174dismantled methodology. These are expressly discarded:175176- ❌ **The "10x programmer".** Origin: **Sackman, Erikson and Grant, "Exploratory Experimental Studies177 Comparing Online and Offline Programming Performance", *CACM* 11(1):3-11, Jan-1968**. Methodological178 refutation: **Lutz Prechelt, "The 28:1 Grant/Sackman legend is misleading, or: How large is179 interpersonal variation really?", Technical Report 1999-18, Universität Karlsruhe, Dec-1999**180 (`page.mi.fu-berlin.de/prechelt/Biblio/varianceTR.pdf`). Flaws in the original: **12 subjects**,181 trivial tasks of a mathematical nature, and **a mix of low-level and high-level language182 programmers in the same group**, which inflates the best/worst ratio. Prechelt reanalyses a much183 larger dataset and argues that **comparing the best with the worst is the wrong comparison**: against184 the median of the worst quartile versus that of the best, the ratio **rarely exceeds 4**, and the185 standard deviation/mean ratio is around **0.5**. **Permitted use**: "there is substantial186 interpersonal variation, of the order of 2-4x depending on the task, with overlapping distributions". **Forbidden use**:187 "10x", and above all **using it to justify pay, staffing or redundancies**.188- ❌ **"An interruption costs 23 minutes and 15 seconds".** **The paper that backs it does not189 exist.** The paper that gets cited, **Mark, González and Harris, "No Task Left Behind? Examining the190 Nature of Fragmented Work", CHI '05**, measures the **probability of resuming a task on the same day**191 and **does not contain that figure**. The other one usually cited, *"The Cost of Interrupted Work: More192 Speed and Stress"*, finds that **interrupted subjects completed the tasks faster**,193 with more stress. The figure comes from **a 2006 interview** with Gloria Mark, not from a peer-reviewed194 article. **It is not written, not even with caveats.** What is defensible without a figure: fragmentation has a cost195 and **the argument for protecting blocks of work does not need an invented number**.196- ❌ **"70 % of the variance in team engagement is determined by the manager"** (Gallup, *State197 of the American Manager*, 2015). **Proprietary analysis, not peer reviewed, with no raw data198 published**, and with two interpretation problems that its diffusion ignores: it refers to variance199 **between teams** (it is not an individual-level claim), and the "manager effect" absorbs everything200 that clusters at team level (function, location, load, staffing). **It is not used as data to201 justify a decision.**202- ❌ **"People leave managers, not companies"** (*First, Break All the Rules*, Gallup). There is a published203 refutation from **Culture Amp (2017)** over 175 teams: people do leave bad managers, but **it is not204 the first reason**, and in bad organisations **having a good or bad manager barely changes the decision to205 leave**. The weak version holds ("the manager's quality is *one* of the factors, conditional on206 the organisation being decent"); the slogan does not.207- ❌ **Turnover percentages attributed to a "bad manager"** and **the cost of a bad hire208 expressed as a multiple of salary** (§ `technical-hiring-standards`): discarded for the same209 reason, there in detail.210211**General rule**: **a figure with no primary study, year, sample and method does not go into a document,212a slide or a budget request.** If the argument only stands with the figure, the213argument was bad.214215## 3. Structure and conventions216217### 3.1 Career tracks: what actually changes218219**What changes when you move up a level is not "the level": it is three observable variables.**220221| Variable | Senior | Staff | Principal |222|---|---|---|---|223| **Scope of impact** | Their team and the system they answer for | Several teams, or a critical area the organisation depends on | The engineering organisation or a cross-cutting capability |224| **Type of decision** | How it is implemented; chooses between known options | Which options exist; defines the problem and the interfaces | What is stopped; irreversible decisions and trade-offs between areas |225| **Time horizon** | Weeks to a quarter | Quarters to a year | One to several years |226| **Required evidence** | Systems delivered and operable | Other people's designs improved, decisions recorded, people levelled up | Changes of direction the organisation followed and that can be traced |227228Hard rules:2291. **The ladder is published.** A ladder only the manager knows is not a ladder, it is a retroactive230 excuse. Public references exist so as not to start from zero — **Rent the Runway (2015, with231 Camille Fournier as CTO), Dropbox, CircleCI, Square, Kickstarter**, aggregated at232 `progression.fyi` and `github.com/bmoeskau/engineering-ladders`. **They are adapted, not copied**: someone233 else's ladder describes someone else's organisation.2342. **Real parity between tracks.** The individual contributor track and the management one have **the same235 level ceiling, the same salary band and the same presence in decision-making forums**. Without those236 three, the technical track is decorative and everyone knows it within six months.2373. **Managing is not a promotion, it is a change of job.** The work, the measure and the skill are238 different. Operational corollary: **there is a declared way back**, and using it is not a failure239 recordable in the evaluation.2404. **The level is assigned on written evidence**, not on seniority and not on likeability. If the evidence cannot241 be written down with dated facts, there is no level.2425. **No level is defined by a technology.** "Kubernetes senior" is not a level, it is a243 dependency on a tool with an expiry date.244245### 3.2 Staff+ archetypes: what they are really for246247They serve one purpose, and that is why they are here: **avoiding assigning work the person cannot248do from where they are.** Taxonomy and caveats in §2.249250| Archetype | Work that belongs to them | Work that destroys their impact |251|---|---|---|252| **Tech lead** | Approach and execution of **one** team, alongside the manager | Being the bottleneck for the decisions of another three teams |253| **Architect** | Direction, quality and approach of **one critical area**, with a multi-year horizon | Firefighting outside their area; making their area depend on their presence |254| **Solver** | Going into a hard, bounded problem and coming out with the path resolved | Becoming the permanent owner of what they fixed |255| **Right hand** | Extending an executive's reach in a large organisation | Existing in an organisation that does not need one: it reproduces hierarchy without adding judgement |256257Rules:258- **The archetype is agreed in writing with the person and is reviewed** (by default, every six months or259 when the project changes). A *solver* who gets handed the maintenance of everything they touched260 stops being a *solver* within a quarter.261- **Architect and right hand only appear at a certain scale.** Larson observes them emerging around262 ~100 and ~1,000 engineers respectively. **Creating them earlier manufactures a decision layer with no263 problem to solve.**264- **FORBIDDEN to use them as titles** or as levels on the ladder (§7).265266### 3.3 Decisions: what is decided, who decides and how fast267268**Who decides, by default:**269270| Type of decision | Decided by | Consulted | Recorded in |271|---|---|---|---|272| Implementation inside a module of the team's own | **Whoever implements it** | The PR reviewer | The code and the PR |273| Choice of library or pattern within the team | **The team**, the tech lead arbitrates | — | **The repository's ADR** |274| Contract or interface between two teams | **Both teams, jointly** | The area's architect | **ADR + versioned contract** (`api-design-standards`) |275| A new technology in the organisation, or the retirement of an existing one | **Staff+ level or architecture**, with published criteria | Platform, security, FinOps | **ADR + inventory** |276| A scope or date commitment | **Not this skill** | — | `project-management-standards` |277| Acceptance of a regulatory risk | **Whoever has the formal authority** | — | `grc-compliance-standards` |278279**The ADR is an obligation, not a courtesy.** Minimum rule: **if a technical decision is going to condition the280work of someone who was not in the conversation, it gets written.** Nygard format (§2), in the281affected repository, numbered and not editable once accepted.282283**Speed is set by reversibility, not by perceived importance** (Bezos 2015, verbatim in §2):284285| | One-way door (Type 1) | Two-way door (Type 2) |286|---|---|---|287| Typical examples | Public data model, identity scheme, API format exposed to third parties, lossy migration, choice of a vendor with an expensive exit | Internal library, folder structure, internal log format, test framework |288| Speed | **Slow, deliberate, informed**: design document, alternatives evaluated, formal review | **Fast**, by the person or the small group closest to the problem |289| Level that decides | Higher up, with consultation | The lowest possible |290| Cost of being wrong | High and permanent | Low: the door reopens |291292- **Every ADR carries a `reversibilidad` field with an explicit value.** Without it, the organisation treats293 everything as Type 1 and grinds to a halt — the failure the 2015 letter itself describes.294- **The classification can be challenged, but in writing and with an argument.** "This is irreversible"295 is a falsifiable claim: you ask for the estimated cost of reverting it.296- **A Type 2 decision that has been under discussion for three weeks already cost more than being wrong.** Rule:297 a time box, and if it expires, the person with the responsibility decides and it is recorded.298299### 3.4 Design document and design review300301**The design review is the point in the cycle where a mistake is still cheap.** Afterwards there is302code, dependencies, migrated data and people who have already defended it in public.303304**When a design document is mandatory** (any one is enough):305- The decision crosses a team boundary or creates/changes an interface between teams.306- It is a one-way door (§3.3).307- It touches personal data, authentication, authorisation, cryptography or money.308- The estimated effort exceeds a threshold the team publishes (by default: **two person-weeks**).309- A technology the organisation does not yet operate is introduced.310311**Minimum template. A document without sections 5, 6 and 8 is sent back unreviewed.**312313```markdown314# Diseño: <nombre> Autor: <persona> Estado: borrador|en revisión|aceptado|sustituido3151. Problema # qué falla hoy, con evidencia observable (dato, incidente, ticket)3162. Restricciones # plazo, presupuesto, normativa, compatibilidad, equipo disponible3173. No objetivos # lo que este diseño explícitamente NO resuelve3184. Propuesta # la solución, con el diagrama mínimo que la explique3195. Alternativas descartadas # >=2, con el motivo del descarte. Sin esto no hay diseño, hay preferencia3206. Criterios de aceptación # falsables y medibles: cómo se sabrá que funcionó, y cuándo se mira3217. Impacto # migración, compatibilidad hacia atrás, coste recurrente, operabilidad, seguridad3228. Reversibilidad # Tipo 1 o Tipo 2, y coste estimado de revertir3239. Riesgos abiertos # con dueño324```325326**Review rules:**327- **The problem is reviewed before the solution.** If the reviewers do not agree on the328 problem, discussing the solution is wasted time and it always ends in aesthetics.329- **Mandatory pre-reading, with a deadline** (48 h by default). A meeting where the document is read330 live is a reading meeting, not a review.331- **Comments are labelled as blocking or non-blocking**, just as in code review332 (`code-review-standards`). A reviewer who marks nothing as blocking has approved.333- **The review ends with a decision and a date, not with "let's keep talking".** The334 valid states are: accepted, accepted with listed conditions, rejected with a reason, or postponed with335 a date and with what remains to be found out.336- **The accepted document becomes an ADR or links to one.** A design that leaves no trace of a decision337 gets discussed again in six months.338- **Acceptance criteria with a check date.** They are revisited when due: if they were not met,339 that is information about how the team designs, and it is the cheapest input there is for improving.340341### 3.5 Technical debt: a business decision with a budget342343The technical criteria —what to refactor, with which technique and with which test net— belong to344`refactoring-tech-debt-standards` (**already written**). **Here only the decision to invest and345how it is defended.**3463471. **Debt is recorded like any other work**, in the same backlog, with an owner and with the348 **cost it imposes today** — not with adjectives. Admissible formulation: *"every change in the349 billing module requires touching four places and last quarter it generated 3 of the 7 incidents"*.350 Inadmissible formulation: *"the code is bad"*.3512. **Distinguish deliberate debt from degradation.** Deliberate debt was incurred with a recorded352 decision and a review date (if it does not have one, it was not a decision: it was an oversight).353 Degradation is silent accumulation and is detected by symptom, not by opinion.3543. **An explicit and recurring budget.** A percentage of capacity per cycle is declared,355 published and **protected the way a committed date is protected**. A percentage that gets cancelled in356 the first quarter under pressure was not a budget: it was an intention.3574. **How it is defended to whoever pays — and this is what always fails.** Technical debt is not358 defended as code quality: **it is defended with the cost the business is already paying**, in359 its vocabulary. Arguments that work because they are verifiable:360 - **Lead time**: the same change takes X in this module and Y in the rest.361 - **Reliability**: the proportion of incidents concentrated in the component362 (`incident-management-standards` provides the datum).363 - **Direct cost**: oversized infrastructure, licences, hours of manual operation364 (`finops-standards`).365 - **Risk with a name**: an unsupported dependency, a version with no security patches366 (`vulnerability-management-standards`), regulatory non-compliance (`grc-compliance-standards`).367 - **The option that is lost**: what will not be doable, and when, if it is not touched.3685. **A refactor request with none of those five arguments is rejected**, whoever makes369 it. And it is rejected here, inside engineering, before the business rejects it: **credibility370 is spent once**.3716. **Never "a refactor release" that delivers nothing.** Debt work is delivered in372 increments with an observable effect; if it cannot be split, the problem is the design of the373 intervention, not the calendar.374375### 3.6 The team as a system376377- **A single owner per system.** A system with no named owner is maintained by whoever had the bad luck378 of touching it last. Two owners is zero owners.379- **Bounded load, in the analogical sense of cognitive load** (§2, with its caveat): a team380 answers for as many systems as it can **understand, operate and improve**. An operational and381 falsifiable test that requires no invented metric: *can the team deploy, diagnose in382 production and explain the data model of every system it answers for, without depending on one383 specific person?* If not, the load exceeds the team — the scope is reduced or extraneous load384 is transferred to the platform (`platform-engineering-standards`).385- **Stable teams.** Reorganising teams resets system knowledge and trust386 relationships. **Every reorganisation is justified in writing with the problem it solves and its expected387 cost**, or it is not done.388- **Size**: small enough that everyone knows everyone's work, large enough389 to sustain an on-call rota without burning anyone (`sre-practice-standards` fixes the latter).390 **Adding people to a saturated team adds coordination before it adds capacity**391 (`project-management-standards` §6).392- **Inter-team dependencies: they are managed by interface, not by meeting.** Preference, in order:393 (1) eliminate the dependency, (2) decouple with an interface contract and a simulator, (3) sequence394 with a mutual commitment from the providing team, (4) escalate. **Adding a recurring395 sync meeting is choosing (4) and calling it (3).**396- **The team's structure and the system's converge** (Conway). Practical consequence: **if you397 want a different architecture, you have to change the team boundaries, not just the398 diagram.**399400### 3.7 Technical conflict401402**A technical disagreement with no procedure gets resolved by hierarchy or by exhaustion, and both403produce the worst decision available.** Procedure, in this order and with a deadline:4044051. **Make the disagreement explicit**: each side writes, in one paragraph, what they defend and **what evidence406 would make them change their mind**. Whoever cannot answer the second has no technical position,407 they have a preference — and preferences do not block.4082. **Find the datum**: prototype, measurement, load test, review of a past incident. **With409 a fixed time box agreed beforehand.**4103. **If the datum does not discriminate** (a real tie), **whoever answers for the consequences** decides —411 normally the owner of the affected system — and it is recorded in an ADR with the alternatives and the412 tie declared.4134. **Disagree and commit, as an explicit team rule.** Reference: Amazon's leadership414 principle **"Have Backbone; Disagree and Commit"** (`amazon.jobs`) — the obligation to challenge the415 decision respectfully even when it is uncomfortable, and **full commitment once it is taken**. The term is416 older than Amazon (Intel, *constructive confrontation*). The local rule, which is what applies:417 **once the decision is closed, nobody passively sabotages it or reopens it in corridors**; it is reopened418 only with new evidence and through the same channel.4195. **When it is escalated**: when the disagreement crosses the boundary of two teams and blocks work for more420 than a week, or when it involves a security, legal or personal-data risk. **Escalating is not421 losing**: escalating late is.4226. **A disagreement that repeats is an organisational design problem**, not a people problem: the423 boundaries of responsibility are badly drawn (§3.6).424425### 3.8 The technical leader's work that is not writing code426427**Their value is measured in the team's result, not in their own *output*.** Direct consequences:428if the most productive person on the team is the leader, the team is underused; if the leader429is on the critical path of every delivery, they are a SPOF with holidays.430431Work that is theirs, and that produces leverage:432- **Unblocking**: identifying who is stuck and why, and getting it out of the way. It is the highest-return433 activity of the job and the first one sacrificed when the leader starts programming.434- **Defining interfaces between teams** before two teams build against different435 assumptions (§3.6).436- **Reviewing other people's designs and decisions** (§3.4) — not rewriting them.437- **Writing what the organisation will need to remember**: ADRs, context, criteria.438- **Creating the conditions for someone else to take the decision**, and accepting that they take it differently from how439 the leader would have taken it if the result is acceptable. **A leader who only delegates the decisions that440 would have matched their own has delegated nothing.**441- **Developing people** (§3.9): it is work with a calendar, not a by-product.442443**The opposite trap, and it is just as real: the leader who stops touching the code loses the ability444to judge.** With no contact with the system, other people's estimates become uncheckable, designs445get approved out of trust in the person and not because of the content, and technical debt becomes446a story with no evidence. **Non-negotiable operational minimum:**447- Review code properly and regularly (`code-review-standards`), including changes they do not448 fully understand, which is where you learn where the system is.449- **Be in the on-call rotation** if the team has one, or at least take part in incident450 response (`incident-management-standards`).451- **Take real work, but never on the critical path**: internal tooling, fixes,452 tests, the task nobody wants. **FORBIDDEN to take the quarter's critical feature**:453 it turns them into a bottleneck and leaves the team without the developmental part.454- Arbitration rule between the two traps: **if the team delivers just as well when the leader is455 away for two weeks, the split is correct.** It is the only test that matters, and it can be run456 literally.457458**Code review is not used as an instrument of power.** Vetoed manifestations: blocking a459PR out of aesthetic preference, demanding your own solution with no technical argument, withholding approval460as leverage in a negotiation unrelated to the diff. The criteria for what blocks belong to461`code-review-standards`; **the obligation that the leader does not distort it belongs here**.462463### 3.9 1:1s, feedback and evaluation464465**The 1:1 belongs to the report.** Concrete rules:466- **The agenda is set by whoever reports**, and it exists before the meeting. If there is no agenda, the default467 question is not "how is X going?" but **"what is getting in your way most?"**.468- **FORBIDDEN for it to be a status report**: the state of the work is in the management system469 (`project-management-standards`) and consulting it is the leader's responsibility, not minutes of the470 person's time.471- **A fixed and protected cadence** (weekly or fortnightly by default). **Cancelling it repeatedly472 communicates a priority more clearly than any speech.**473- **Shared notes and commitments with an owner and a date.** A 1:1 with no trace repeats identically for474 a year.475- At least once a quarter, the 1:1 is about **career**: current level, missing evidence,476 archetype (§3.2), what work needs to be sought to get there. With the published ladder in front of you.477478**Feedback:**479- **Specific, dated and about observable behaviour**, not about traits. *"In the review of the payments480 design on 12 May you closed the discussion before Ana presented her alternative"* is481 actionable. *"You are not very collaborative"* is not: what cannot be observed cannot be corrected.482- **Close to the event.** Feedback saved for six months for the annual review is not feedback: it is an483 ambush, and it destroys the credibility of the whole process.484- **No surprises in the evaluation.** Falsifiable rule: **if something appears in the annual review and it is485 the first time the person hears it, the failure is the leader's, and it is recorded as such.**486- Difficult feedback is given in private, in writing as well as verbally, and **with what is expected487 differently and by when**.488489**Why individual engineering productivity metrics are counterproductive** — and this490is the part you have to be able to defend to management with the source in hand:4911. **The published position of the sector's reference research.** **DORA**: the four metrics492 measure **team and system delivery** and must not be used to evaluate the performance of493 individual engineers; applying them to people creates perverse incentives. **SPACE** (Forsgren et494 al., 2021, §2) dismantles two directly applicable myths: that productivity is activity, and495 that it is **only** individual performance; and it establishes that **activity metrics are never used496 in isolation to reward or penalise**.4972. **They are trivial to game and gaming them is rational**: lines, commits, PRs and points are498 free to inflate and the cost of inflating them is paid by someone else (the reviewer, the maintainer).4993. **They measure the visible and punish the valuable**: reviewing, pairing, mentoring, being on call,500 deleting code and preventing the unnecessary from being built leave no trace in any counter.5014. **Engineering work is interdependent**: attributing the result to an individual within502 a system with queues, dependencies and review is an attribution error, not an imprecise503 measurement.5045. **The alternative that is used**: qualitative evaluation **with written and dated evidence** against505 the published ladder (§3.1) — designs, recorded decisions, incidents resolved, people506 developed, systems that work without their author. **It is more work for the leader. That is the507 cost of the job.**508509### 3.10 Incidents: what corresponds to leadership510511The full process belongs to `incident-management-standards`. **Here only the part that only512leadership can do, and which is where the blameless culture breaks in practice:**513- **Protecting the blameless postmortem when there is pressure from above to name someone.** The514 pressure is real and arrives in the form of a reasonable question ("who deployed it?"). The standard515 answer, in the same language as whoever asks: **the system allowed an individual error516 to reach production; that is the defect and that is what gets fixed, because it is the only one that will not517 recur.** Replacing the person leaves the system identical for the next one.518- **Verifiable drafting rule**: **the postmortem contains no proper names as a cause.** If519 "X ran Y" appears, it is rewritten as "the procedure allowed Y to be run without confirmation".520- **Honest contraindication**: blameless **does not mean without responsibility**. Repeated521 negligence, knowingly bypassing controls or acting without authorisation are a conduct matter and are522 handled **outside the postmortem, through another channel and with HR.** Mixing them destroys both.523- **The postmortem's actions get funded.** A postmortem whose actions do not enter the next524 cycle's plan teaches the team that the exercise is theatre, and from then on postmortems525 are written to be filed.526- **The leader publicly owns the team's failure and publicly attributes the success.** It is a527 conduct rule with a verifiable effect: without it, nobody reports a problem early.528529### 3.11 Coding agents: what leadership decides530531The concrete policy —which task is delegated, instruction files, permissions, review of the diff,532attribution— belongs to `ai-agent-workflow-standards`. **Here, three 533534…(truncated)