System Optimization

System Optimization Protocol

Core Directives

1. Eliminate Waste First (Lean): Remove non-value-adding activities before optimizing what remains.
2. Fix the Constraint (ToC): The weakest link sets the throughput ceiling. Find it; subordinate everything else.
3. Stabilize Before Optimizing (Six Sigma): An unstable process cannot be meaningfully improved.
4. Build Quality In — Shift Left (DevOps): Embed quality at the source via automation, types, and linting.
5. Small Steps (Kaizen): Many small validated improvements compound faster than infrequent large redesigns.
6. Decide Late (Lean): Defer irreversible commitments until you have the most information.
7. Manage Flow, Not Utilization (Queueing): Throughput is limited by queues, batch size, and variability — not by how busy each resource is. High utilization creates delay; it does not cure it.

Order of Operations

Apply these steps in order. The most expensive mistake is optimizing or automating something that should have been deleted — work done out of order is work to undo later.

1. Question the requirement. The requirement is the most upstream waste. Strip it to first principles before touching the implementation; the cheapest step is the one that no longer needs to exist. Separate each requirement's obligation (the outcome, evidence, or restriction that must exist) from its mechanism (the specific rights, routes, roles, record types, or protocols it names). Obligations are floors; pinned mechanism is negotiable — audit within the current requirements first to establish the floor, then propose careful requirement edits scored like any change, and gate edits with real external trade-offs as explicit product decisions (see functionality-complexity-tradeoff §1e).
2. Probe deletion. Try removing the step, file, stage, or component behind a reversible branch, feature flag, dry run, or narrow rollout. Restore anything proven load-bearing; keep only deletions backed by evidence.
3. Simplify what remains. Only optimize parts that survived deletion.
4. Speed it up. Parallelize, cache, batch — but only after simplification, and only within the limits parallelism actually returns (§7).
5. Automate last. Automating an unnecessary or unsimplified process locks in the waste.

1. Scan Layers

2. Waste Scan (Lean — Muda / Mura / Muri)

Lean names three families of waste. Muda is non-value-adding work; the TIMWOODS table below enumerates its eight forms. But Muda is usually manufactured by two deeper causes — attack these at the source first:

Muda (TIMWOODS):

Flag every step where wait time > cycle time — that is queued inventory (§4).

3. Constraint Identification (ToC — 5 Steps)

1. Identify the current constraint: the step or shared resource with the lowest effective throughput or largest queue.
2. Exploit — maximize its output without adding resources.
3. Subordinate — ensure upstream steps don't feed it faster than it can consume.
4. Elevate — if still a bottleneck, invest in capacity.
5. Repeat — re-measure after each change; the constraint may move.

Most constraints are policy, not physical capacity (Goldratt): a batching rule, a mandatory approval, an unlimited-intake WIP policy, a "single reviewer" convention. Check for a policy constraint before buying capacity — elevating a policy costs nothing but agreement.

4. Flow Economics (Queueing)

Queues, not busy resources, set delivery time. Three laws make flow measurable and turn "improve flow" into concrete moves.

• Little's Law — for any stable queue, cycle time = WIP / throughput. To ship faster without adding capacity, cut work-in-progress: halving in-flight work halves average cycle time at the same throughput. This is why WIP limits beat exhortations to "go faster."

• Utilization curve (Kingman) — queue time rises non-linearly with utilization ρ, roughly wait ∝ ρ / (1 − ρ); past ~80% it explodes:

  ρ (utilization)	0.5	0.8	0.9	0.95
  wait ≈ (× service time)	1×	4×	9×	19×

  A resource pinned at 95% is not "efficient" — it is a delay generator (this is Muri, §2). Size shared resources (CI runners, reviewers, on-call) for headroom, not full utilization.

• Batch size — large batches inflate cycle time, delay feedback, and raise risk per release. Halving batch size roughly halves queue time and localizes failure. Small PRs, trunk-based flow, and incremental deploys are batch-size controls, not stylistic preferences.

5. Measure the System

You cannot optimize what you cannot measure, and you cannot measure an improvement you cannot separate from noise. "Go and see" the real signal — logs, queues, traces, timings — rather than optimizing from assumption.

Delivery & flow metrics

Track a small, stable set; a regression in any of them is a bug (§7).

Flow efficiency is the highest-yield diagnostic: map the value stream, split each step into work time vs wait time, and the wait total is your inventory. Low flow efficiency means the fix is removing queues, not speeding up steps.

Error budget (spend or stabilize) turns the reliability metrics into a decision rule that resolves the tension between Directives 3 and 7. Set an explicit failure tolerance (an availability or change-fail-rate target); the gap between target and actual is the budget. While budget remains, spend it — ship faster, take risk, raise deploy frequency (Directive 7). Once it is exhausted, freeze feature risk and stabilize until it recovers (Directive 3). This makes the speed-vs-reliability trade-off explicit rather than political. The pipeline gates that measure and enforce these signals live in ci-cd-reliability-architecture; the budget-spend policy is an optimization decision and lives here.

Variation: common vs special cause (Six Sigma)

Before reacting to any metric move, classify the variation:

Reacting to common-cause noise as if it were signal (tampering / over-adjustment) adds variance — the classic Deming funnel error. Establish a baseline band first; treat only out-of-band moves as regressions. This is what "stabilize before optimizing" (Core Directive 3) demands operationally, and it is why flaky tests are measurement-blockers, not nuisances.

Run structured improvement as DMAIC — Define, Measure, Analyze, Improve, Control — the Six Sigma loop that complements Kaizen/PDCA (§11): PDCA drives many small reversible bets; DMAIC governs a defect or variance that needs a controlled, evidence-gated fix locked in by a Control step so it cannot regress.

6. Diagnostic Reasoning Chain

For identifying constraints, waste, and root causes in operational analysis:

• First Principles: Strip legacy assumptions. Rebuild from observed facts, current requirements, and explicit constraints.
• 5 Whys: Ladder from the observed symptom to a systemic root cause. Stop at the first cause you can change in the system, not the instance.
• Analogical Reasoning: Apply patterns from other domains to local problems.
• Constraint Removal: Imagine the ideal solution with no legacy debt before committing.
• Inversion: "How do I guarantee failure?" — then build guardrails against it.
• Simple Functional Refactor Over Complex Technical Solution: Prefer reasonable functional or UX changes (clear a cache, reorder steps, show a waiting state) before technical solutions (polling, retries, new abstractions). Technical complexity carries hidden costs — validate that simple alternatives don't solve it first.
• Pre-mortem: Assume the fix already failed; work backward to find the oversight.
• Side Effect Audit: When eliminating redundancy, trace all downstream paths that depended on the original behavior.
• Pattern Parity: Do not let divergent legacy patterns coexist indefinitely with a newly established standard. Temporary migration overlap needs an owner, scope, and removal condition.

7. CI/CD

• Bottleneck first (ToC): Optimize the measured constraint before adjacent stages.
• Parallelize safely: Independent tests, builds, and linting can run concurrently when artifacts, caches, rate limits, and shared environments are isolated. Keep dependent or resource-contentious stages sequential. Speedup is capped by the serial fraction (Amdahl); past a point, coordination and contention cost make more parallelism slower (Universal Scalability Law). Measure before widening a fan-out.
• Idempotent environments: Deployment state must be reproducible from source control.
• Shift-left gates: Linting and unit tests run before integration tests.
• Measure: Track build duration, failure rate, and flakiness. Regressions are bugs.
• Zero-downtime secret rotation: Always create the new credential first, apply it to all environments, then delete the old one. Never delete before applying — the gap causes downtime and broken deployments.

8. Developer Workflow

• Small PRs: Faster merge, smaller blast radius, better review quality. Large PRs are inventory and oversized batches (§4).
• Short feedback loops: Fast local test results reduce context-switching cost.
• Limit work-in-progress: Cap in-flight work per person and per stage; a lower WIP ceiling shortens cycle time more reliably than working faster (§4).
• Eliminate toil: Recurring manual tasks should be removed, simplified, or automated only after the task is necessary, stable, and cheaper to automate than to perform manually. For where the work should ultimately live, see push-out.
• One-piece flow: Work moves design → build → review → deploy without sitting idle.

9. Testing

• Detection distance: Bugs caught closest to their source are usually cheapest. Use the earliest check that can actually detect the defect; some boundary and workflow failures require integration or e2e tests.
• Flakiness is a defect: A flaky test erodes trust and masks real failures; it is also a measurement-blocker (§5).
• Confidence over coverage: Optimize for critical-path confidence, not line percentages.
• Shift-left security: Static analysis, dependency audits, and secret scanning run in CI.

10. Documentation

• Executable specs over text: Tests and self-documenting code are living documentation.
• JBGE: Minimal, concise, audience-specific. Documents with no owner, current reader, legal/compliance need, or incident value should be archived or deleted after confirming they are not load-bearing.
• ADRs: Document why, not what. Prevents future rework from revisiting settled decisions.

11. Continuous Improvement (Kaizen / PDCA)

• Every optimization is a hypothesis — validate before declaring permanent.
• After resolving a bottleneck, explicitly identify the current constraint before the next cycle.
• When several improvements compete and none is the constraint, sequence by cost of delay ÷ effort (WSJF): the change whose delay is most expensive per unit of work goes first. The constraint (§3) still outranks all of them.
• Use the four axes (D, K, P, n) from structural-simplification to compare per-axis deltas before and after each improvement.

Litmus Test: If a change worsens any complexity axis (D, K, P, n) from structural-simplification without measurably improving flow, reliability, cost, or another complexity axis, it is not an optimization.

12. Output Contract

When applying this skill, emit a coder-facing optimization decision record:

Scope:          <repo / pipeline / workflow / module / time window>
Constraint:     <measured bottleneck or Not measured + reason>
Decision:       Delete | Simplify | Stabilize | Optimize | Automate | Defer | Reject
Flow signal:    <WIP / cycle time / utilization / flow efficiency / variation class, or Not measured>
Evidence:       <metric, log, queue, test result, file path, or user workflow checked>
Next action:    <one concrete edit, experiment, measurement, or owner question>
Verification:   <before/after metric, command, CI run, or Not run + reason>
Residual risk:  <risk left for judgment, rollout, or future measurement>

13. See also

• structural-simplification — per-axis complexity comparison used for Kaizen Δ scoring.
• defect-shift-left — placement of each error detection at the earliest possible stage (operationalizes "Build Quality In").
• ci-cd-reliability-architecture — pipeline-reliability prerequisites that stabilize a process before optimization; owns the delivery gates this skill only measures.
• push-out — where recurring operational work should live once flow is stable.