HLD interview
Resolve architectural direction before prose. This skill runs in the lead agent's context so
questions remain connected to the approved product decisions; never delegate the conversation.
The one exception is discussing a single question the user is unsure about: that runs in an
isolated discussion agent or a side chat, under this skill's read boundary and design-conversation
rules, so its reading never fills the lead's context.
Read boundary
Read only workflow-supplied requirements, decision artifacts, revision feedback, project guidance,
maintained codebase maps, and source files whose detail is necessary for a material decision the
maps do not settle. Do not scan application source broadly or search outside the repository.
Architecture Grill
Ask only decisions that change the approved direction or what a downstream LLD must respect:
- system and repository ownership;
- service, module, and trust boundaries;
- source of truth and major data lifecycle;
- cross-repository control or data flow;
- authentication, authorization, privacy, and sensitive-data exposure;
- failure, recovery, idempotency, concurrency, and relevant execution modes;
- compatibility, migration, rollout, backout, and material operational trade-offs.
Do not ask about PRD scope already settled. Do not ask for endpoint fields, table columns, class or
file names, library syntax, task breakdowns, or other LLD choices. Recommend the simplest direction
grounded in the approved requirement and existing architecture.
Infrastructure questions follow the workspace, not a checklist. When your instructions supply
an infrastructure profile, its level decides:
- none — the workspace defines no infrastructure: ask no deployment, scaling-topology or
platform questions. Scale is answered inside the application design (caching, async work,
pagination) only when a non-functional target needs it.
- containers — ask only the runtime and configuration decisions this feature changes: a new
service or process, ports, config and secrets, resource limits.
- iac — the workspace holds infrastructure as code: read those repositories and ask every
material infrastructure decision this feature changes or depends on — scaling (instances and
autoscaling triggers), data-store capacity and topology, queues and topics, caching, networking
and ingress, config and secrets, environments and rollout, observability (SLOs, alerts), and
cost. Ground each recommendation in what the infrastructure already provisions and in the
requirement's non-functional targets.
At every level, a feature that changes and depends on no infrastructure asks none. Without a
profile, ask infrastructure questions only when the repositories you read define infrastructure.
Design conversation
Architecture is decided with the user, not announced to them.
- For every genuine trade-off, offer the recommendation plus up to two real
alternatives
({label, tradeoff}) with their concrete consequences — cost, coupling, failure behavior,
migration risk. Never invent a straw option to make the recommendation look good.
- When the queue asks anything, end it with one
kind: "direction" question. Its proposal (up
to 150 words) summarises the intended architecture — owners, boundaries, main flow, and the key
trade-offs taken, assuming each recommendation is accepted — so the user confirms, challenges,
or discusses the whole shape before prose is written. Once the user has answered anything, the
queue cannot close until a direction is confirmed no earlier than their last answer; ask a
fresh direction (new ID) only when a later answer changes it. When there is nothing to ask,
ask no direction either.
- When the user is unsure or asks back, explain the reasoning, compare options, and answer from
the requirements, maintained maps, and the specific source that settles it. Say what you
checked. If a fact belongs to someone else, offer a working assumption or deferral.
- Invite the user's own concerns; treat a
user-raised point as a constraint to reconcile.
Answer audit
A saved answer is not automatically complete.
- Compare it with every
must_resolve fact attached to its question.
- For
source: user-answer, treat only answer as confirmed; unaccepted proposal text is not a
decision. chosen-alternative and accepted-proposal are confirmed; assumption is a working
answer to flag as unverified; deferred settles nothing; user-raised is the user's own input.
Build on a discussion summary instead of re-asking what was explored.
- A deferred blocking decision must be re-asked with sharper options; a deferred non-blocking one
goes to LLD.
- Compare decisions for conflicting ownership, authority, state boundaries, timing, failure
behavior, or execution-mode assumptions.
- Follow up only on the missing or conflicting fact and use a new stable ID.
- Treat concrete product rules from the PRD as constraints, not questions to reopen.
- A user may explicitly defer a non-blocking implementation choice to LLD. Preserve that answer;
never describe a blocking architectural decision as safe to defer.
Queue contract
Write the workflow-supplied queue as schema_version: 2 with questions and audit. Each question
has id, title, question, why, proposal, and one to four short must_resolve facts; it may
add alternatives and kind (decision by default, or direction).
audit.unresolved and audit.contradictions name why another round exists. Both arrays are empty
only when the queue is empty and the architecture is clear enough to write. Run the supplied
validator before completing.
Fewest questions. There is no minimum, and an empty queue on the first pass is a normal
result. A question earns its place only when the approved requirement, the decision log, the code and the docs do not settle it and reasonable architects
would choose differently at a cost that is high to change after approval. A point with one sensible answer is not a
question: record it under settled when a source states it, otherwise design it and let the
document's review check it. Put every material
gap in one queue rather than spreading it over rounds.
One decision log, no repeats. Your instructions may supply the feature's decision log (every
earlier stage's decisions) and this stage's decisions. Give every question a short kebab-case
topic naming what it decides. Never ask a topic already decided in either; to refine a recorded
answer, ask at most one follow-up with follow_up_of: <that decision's id>. Rephrasing a decided
question under a new id is still a repeat.
Decide from evidence first. Before asking, check the requirement, approved documents, code and
docs. A point they already settle is not a question: record it under the queue's settled array
as {topic, title, decision, source}, where source is the file that settles it
(path, path#section or path:line). The engine records it as an evidence-settled decision.
Interview presentation
When the workflow serves a question batch, collect it through the requested native popup. If the
host limits questions per popup, collect consecutive chunks without analyzing or recording between
them. Then discuss the items the user was unsure about or asked about, take any point they raise,
and record the complete round atomically.
Standards
- One material decision per question; one question mark.
- State why it matters in one short sentence.
- Prefer a concrete recommendation with real alternatives over an abstract menu.
- Never ask questions derived from HLD headings.
- Never draft or edit the HLD in this skill.
- Keep each turn short and let the user steer; no essays between rounds.
When the workflow supplies validator feedback naming defects in an existing queue, fix only those
defects in that file and validate again; do not rebuild the queue or re-read sources.
Safety
Write only the workflow-supplied question queue. The engine owns durable interview state and the
cumulative decision artifact.
Output contract
Return summary: one short sentence describing the prepared architecture question queue or that
the architecture audit is clear.
1---2name: hld-interview3description: Clarify material architecture decisions before an HLD is finalized. Generate feature-specific technical Grill questions, audit prior answers and contradictions, and collect grouped answers without drafting design prose.4allowed-tools: Read, Grep, Glob, Write, Bash5tags: [sdlc, design, hld]6metadata:7 source: https://skillmd.com/skills/keyvaluesoftwaresystems/hld-interview8---910# HLD interview1112Resolve architectural direction before prose. This skill runs in the lead agent's context so13questions remain connected to the approved product decisions; never delegate the conversation.14The one exception is discussing a single question the user is unsure about: that runs in an15isolated discussion agent or a side chat, under this skill's read boundary and design-conversation16rules, so its reading never fills the lead's context.1718## Read boundary1920Read only workflow-supplied requirements, decision artifacts, revision feedback, project guidance,21maintained codebase maps, and source files whose detail is necessary for a material decision the22maps do not settle. Do not scan application source broadly or search outside the repository.2324## Architecture Grill2526Ask only decisions that change the approved direction or what a downstream LLD must respect:2728- system and repository ownership;29- service, module, and trust boundaries;30- source of truth and major data lifecycle;31- cross-repository control or data flow;32- authentication, authorization, privacy, and sensitive-data exposure;33- failure, recovery, idempotency, concurrency, and relevant execution modes;34- compatibility, migration, rollout, backout, and material operational trade-offs.3536Do not ask about PRD scope already settled. Do not ask for endpoint fields, table columns, class or37file names, library syntax, task breakdowns, or other LLD choices. Recommend the simplest direction38grounded in the approved requirement and existing architecture.3940**Infrastructure questions follow the workspace, not a checklist.** When your instructions supply41an infrastructure profile, its level decides:4243- **none** — the workspace defines no infrastructure: ask no deployment, scaling-topology or44 platform questions. Scale is answered inside the application design (caching, async work,45 pagination) only when a non-functional target needs it.46- **containers** — ask only the runtime and configuration decisions this feature changes: a new47 service or process, ports, config and secrets, resource limits.48- **iac** — the workspace holds infrastructure as code: read those repositories and ask every49 material infrastructure decision this feature changes or depends on — scaling (instances and50 autoscaling triggers), data-store capacity and topology, queues and topics, caching, networking51 and ingress, config and secrets, environments and rollout, observability (SLOs, alerts), and52 cost. Ground each recommendation in what the infrastructure already provisions and in the53 requirement's non-functional targets.5455At every level, a feature that changes and depends on no infrastructure asks none. Without a56profile, ask infrastructure questions only when the repositories you read define infrastructure.5758## Design conversation5960Architecture is decided with the user, not announced to them.6162- For every genuine trade-off, offer the recommendation plus up to two real `alternatives`63 (`{label, tradeoff}`) with their concrete consequences — cost, coupling, failure behavior,64 migration risk. Never invent a straw option to make the recommendation look good.65- When the queue asks anything, end it with one `kind: "direction"` question. Its `proposal` (up66 to 150 words) summarises the intended architecture — owners, boundaries, main flow, and the key67 trade-offs taken, assuming each recommendation is accepted — so the user confirms, challenges,68 or discusses the whole shape before prose is written. Once the user has answered anything, the69 queue cannot close until a direction is confirmed no earlier than their last answer; ask a70 fresh direction (new ID) only when a later answer changes it. When there is nothing to ask,71 ask no direction either.72- When the user is unsure or asks back, explain the reasoning, compare options, and answer from73 the requirements, maintained maps, and the specific source that settles it. Say what you74 checked. If a fact belongs to someone else, offer a working assumption or deferral.75- Invite the user's own concerns; treat a `user-raised` point as a constraint to reconcile.7677## Answer audit7879A saved answer is not automatically complete.8081- Compare it with every `must_resolve` fact attached to its question.82- For `source: user-answer`, treat only `answer` as confirmed; unaccepted proposal text is not a83 decision. `chosen-alternative` and `accepted-proposal` are confirmed; `assumption` is a working84 answer to flag as unverified; `deferred` settles nothing; `user-raised` is the user's own input.85 Build on a `discussion` summary instead of re-asking what was explored.86- A deferred blocking decision must be re-asked with sharper options; a deferred non-blocking one87 goes to LLD.88- Compare decisions for conflicting ownership, authority, state boundaries, timing, failure89 behavior, or execution-mode assumptions.90- Follow up only on the missing or conflicting fact and use a new stable ID.91- Treat concrete product rules from the PRD as constraints, not questions to reopen.92- A user may explicitly defer a non-blocking implementation choice to LLD. Preserve that answer;93 never describe a blocking architectural decision as safe to defer.9495## Queue contract9697Write the workflow-supplied queue as `schema_version: 2` with `questions` and `audit`. Each question98has `id`, `title`, `question`, `why`, `proposal`, and one to four short `must_resolve` facts; it may99add `alternatives` and `kind` (`decision` by default, or `direction`).100`audit.unresolved` and `audit.contradictions` name why another round exists. Both arrays are empty101only when the queue is empty and the architecture is clear enough to write. Run the supplied102validator before completing.103104**Fewest questions.** There is no minimum, and an empty queue on the first pass is a normal105result. A question earns its place only when the approved requirement, the decision log, the code and the docs do not settle it and reasonable architects106would choose differently at a cost that is high to change after approval. A point with one sensible answer is not a107question: record it under `settled` when a source states it, otherwise design it and let the108document's review check it. Put every material109gap in one queue rather than spreading it over rounds.110111**One decision log, no repeats.** Your instructions may supply the feature's decision log (every112earlier stage's decisions) and this stage's decisions. Give every question a short kebab-case113`topic` naming what it decides. Never ask a topic already decided in either; to refine a recorded114answer, ask at most one follow-up with `follow_up_of: <that decision's id>`. Rephrasing a decided115question under a new id is still a repeat.116117**Decide from evidence first.** Before asking, check the requirement, approved documents, code and118docs. A point they already settle is not a question: record it under the queue's `settled` array119as `{topic, title, decision, source}`, where `source` is the file that settles it120(`path`, `path#section` or `path:line`). The engine records it as an evidence-settled decision.121122## Interview presentation123124When the workflow serves a question batch, collect it through the requested native popup. If the125host limits questions per popup, collect consecutive chunks without analyzing or recording between126them. Then discuss the items the user was unsure about or asked about, take any point they raise,127and record the complete round atomically.128129## Standards130131- One material decision per question; one question mark.132- State why it matters in one short sentence.133- Prefer a concrete recommendation with real alternatives over an abstract menu.134- Never ask questions derived from HLD headings.135- Never draft or edit the HLD in this skill.136- Keep each turn short and let the user steer; no essays between rounds.137138When the workflow supplies validator feedback naming defects in an existing queue, fix only those139defects in that file and validate again; do not rebuild the queue or re-read sources.140141## Safety142143Write only the workflow-supplied question queue. The engine owns durable interview state and the144cumulative decision artifact.145146## Output contract147148Return `summary`: one short sentence describing the prepared architecture question queue or that149the architecture audit is clear.