optimize

name	optimize
description	Use when something works correctly but is too slow, too memory-heavy, or too resource-intensive. Measures baselines, implements fixes, iterates until a target is reached. Keywords: optimize, performance, profiling, bottleneck, latency, throughput, memory, benchmark.
argument-hint	<component or area to optimize>
disable-model-invocation	false
user-invocable	true
allowed-tools	Read, Grep, Glob, Bash(git:*), Bash(npm:*), Bash(npx:*), Write, Edit
context	fork

Optimize: Measurement-Driven Performance Improvement

You are running the Optimize workflow. The goal is to make $ARGUMENTS faster, lighter, or more efficient — without breaking correctness. Every decision is grounded in measured data, not intuition.

When to Use

• A feature or system is functionally correct but fails performance targets (latency, throughput, memory, CPU)
• Profiling results exist and need to be acted on
• A regression was introduced and the cause is unknown
• You want to establish a baseline before a risky refactor
• You want to iterate toward a specific measurable target (e.g., "p99 latency under 200ms")

When NOT to Use

• The feature is not yet correct — fix correctness first (use /tracer)
• There is no measurable target — optimization without a goal is speculative and risky
• The "performance problem" is unconfirmed by measurement — profile first before changing code

Overview

Optimize works in 5 phases. Each phase produces a committed checkpoint before the next begins.

Baseline (measure current state)
  -> Profile (locate hot paths)
    -> Identify bottleneck (single highest-leverage target)
      -> Implement fix (one change at a time)
        -> Measure delta (confirm improvement, no regression)
          -> Iterate (repeat until target reached or returns diminish)

---

Phase 1: Establish Baseline

1a. Define the Target

If $ARGUMENTS is vague or no performance target was stated, ask one clarifying question:

• What is the measurable target? (e.g., "p99 latency < 200ms", "memory under 50MB", "throughput > 1000 req/s")
• What is the current observed behavior that prompted this?

Do not proceed without a concrete, measurable target. Optimization without a target is refactoring in disguise.

1b. Run the Baseline Measurement

Choose the measurement method appropriate for the project:

• Benchmarks: Run the existing benchmark suite (npm run bench, cargo bench, pytest --benchmark, etc.)
• Load test: If no bench suite exists, write or invoke a representative load script
• Profiling snapshot: If the target is memory or CPU, take a profiling snapshot of the current state
• Manual timing: For scripted or CLI tools, measure wall-clock time with time or equivalent

Record the baseline numbers precisely:

## Baseline: [component/area]

**Date**: [date]
**Target**: [measurable target]
**Current state**:
- p50 latency: Xms
- p99 latency: Xms
- Throughput: X req/s
- Memory: XMB RSS
- [other metrics relevant to target]

**Measurement method**: [benchmark/load-test/profiler/time]
**Commit**: [git hash of measured commit]

1c. Commit the Baseline

git add [benchmark or measurement files]
git commit -m "$(cat <<'EOF'
optimize: establish baseline for [component]

Measured current performance before any changes.
Target: [measurable target]
Baseline: [key metric and value]

Co-Authored-By: Claude Opus 4.6 <[e-mail]>
EOF
)"

---

Phase 2: Profile

2a. Choose a Profiler

Select the appropriate profiler for the language and environment:

Language	CPU profiling	Memory profiling
Node.js	--cpu-prof, clinic flame	--heap-prof, clinic heap
Python	cProfile, py-spy, memray	tracemalloc, memray
Rust	perf, cargo flamegraph	heaptrack, valgrind --massif
Go	pprof (cpu + mem)	pprof
JVM	async-profiler, JFR	JVM heap dump + MAT
C/C++	perf, gprof, Instruments	valgrind --massif, Instruments

If none of the above apply or no profiler is available, proceed to manual code analysis (Phase 3) using Grep and Read.

2b. Run Under Load

Run the profiler against a representative workload — the same workload used for the baseline if possible. Avoid synthetic microbenchmarks unless the target is a specific function in isolation.

2c. Produce a Flame Graph or Hot Path Report

Extract the top consumers from the profiler output:

## Profile Results: [component]

**Profiler**: [name and version]
**Workload**: [what was profiled]

**Top CPU consumers:**
1. `[function/module]` — X% of samples — [file:line]
2. `[function/module]` — X% of samples — [file:line]
3. `[function/module]` — X% of samples — [file:line]

**Top memory allocators (if applicable):**
1. `[allocation site]` — XMB — [file:line]
2. `[allocation site]` — XMB — [file:line]

**Observations:**
- [Any non-obvious finding: e.g., "98% of CPU is in JSON parsing, not the DB query"]

---

Phase 3: Identify Bottleneck

3a. Select One Target

From the profile results, select one bottleneck to address in this iteration. Criteria:

1. Highest leverage: Addresses the largest share of measured cost
2. Feasibility: Can be changed without breaking the interface contract
3. Reversibility: The change can be reverted cleanly if it causes regressions

State the bottleneck explicitly:

## Bottleneck: [iteration N]

**Target**: [function/module/pattern]
**Rationale**: [why this is the highest-leverage target — cite profiler numbers]
**Hypothesis**: [what change is expected to help and why]
**Risk**: [what could break or regress]
**Not targeting**: [other hot paths deferred and why]

3b. Verify the Root Cause

Before implementing, read the hot path code to confirm the hypothesis:

• Use Read and Grep to find the implementation
• Check if there are existing tests covering this path
• Identify callers that will be affected by a change

Do not proceed with a fix until the root cause is confirmed in code — not just in profiler output.

---

Phase 4: Implement Fix

4a. One Change at a Time

Implement one targeted change per iteration. Do not bundle multiple optimizations — each change must be independently measurable. Common patterns:

• Caching: Memoize or cache expensive repeated computations
• Batching: Replace N individual operations with one batched operation
• Lazy evaluation: Defer expensive work until it is actually needed
• Algorithm substitution: Replace O(n²) with O(n log n) or better
• Allocation reduction: Reuse buffers, avoid intermediate copies
• Parallelism: Move independent work off the critical path
• Data structure substitution: Swap a list for a map, a map for a set, etc.
• I/O reduction: Reduce round trips, increase payload size, use streaming

4b. Preserve Correctness

After implementing, run the existing test suite. If tests fail, fix the implementation — not the tests. The optimization is invalid if it changes observable behavior.

[project test command]

4c. Commit the Fix

git add [modified files]
git commit -m "$(cat <<'EOF'
optimize: [brief description of change] in [component]

Bottleneck: [what was identified in Phase 3]
Fix: [what was changed and why it should help]
Hypothesis: [expected improvement]

Tests: passing
No behavior change.

Co-Authored-By: Claude Opus 4.6 <[e-mail]>
EOF
)"

---

Phase 5: Measure Delta

5a. Re-Run the Benchmark

Run the same measurement method used in Phase 1 against the same workload. Record results in the same format:

## Delta: Iteration N

**Commit**: [git hash of fix]
**Fix applied**: [description]

**Before**: [key metric and value from baseline or prior iteration]
**After**: [key metric and value]
**Delta**: [improvement as absolute value and percentage]

**Regression check**: [any other metric that got worse]
**Tests**: [passing / failing — if failing, revert the fix]

5b. Evaluate

• Improvement confirmed and target reached → proceed to Phase 6 (Report)
• Improvement confirmed but target not yet reached → return to Phase 3 (next bottleneck)
• No measurable improvement → revert the fix, return to Phase 3 with a different hypothesis
• Regression → revert the fix immediately, record what broke, return to Phase 3

# If reverting:
git revert HEAD --no-edit

---

Phase 6: Report

6a. Optimization Summary

Produce a final report covering all iterations:

## Optimization Report: [component/area]

**Target**: [measurable target stated in Phase 1]
**Outcome**: [target reached / partially reached / not reached — explain]

### Iterations

| # | Bottleneck | Fix | Before | After | Delta |
|---|-----------|-----|--------|-------|-------|
| 1 | [target] | [change] | [metric] | [metric] | [%] |
| 2 | [target] | [change] | [metric] | [metric] | [%] |
| N | ... | ... | ... | ... | ... |

### Final State vs Baseline

| Metric | Baseline | Final | Change |
|--------|---------|-------|--------|
| [metric 1] | X | Y | +/-Z% |
| [metric 2] | X | Y | +/-Z% |

### What Was Not Addressed

- [Deferred bottleneck]: [why it was not worth addressing]
- [Deferred concern]: [why it was out of scope]

### Known Risks

- [Risk introduced by the optimization]
- [Workload assumptions that may not hold in production]

6b. Create Follow-Up Tasks

If follow-up work was identified:

Create follow-up tasks using your preferred task tracking approach. Each task should include the title, priority, and context from the optimization work. If no tracker is configured, write follow-up tasks to TODO.md in the project root.

---

Guidelines

1. Measure before changing. Every optimization decision must be grounded in profiler data or benchmark numbers, not intuition.
2. One change per iteration. Bundling changes makes it impossible to know what helped or hurt.
3. Preserve correctness. Run tests after every fix. Revert immediately on failure.
4. Revert freely. If a fix shows no improvement, revert it. Dead optimizations add complexity with no benefit.
5. Know when to stop. If the target is reached, stop. If iterations show diminishing returns (< 5% improvement per iteration), report and stop — further work is unlikely to yield results.
6. Don't optimize what you haven't profiled. Hot path intuition is frequently wrong. The profiler is the authority.
7. Document the baseline. Future sessions need the baseline commit hash and numbers to know whether a regression was introduced.
8. Commit each phase. Phase 1 (baseline), Phase 4 (each fix), and Phase 5 reverts should all be committed. This gives git bisect a clean signal.

See Also

• /tracer — iterative correctness-first implementation (use before /optimize if the feature is incomplete)
• /test-strategy — ensure tests exist before optimizing (correctness is a precondition)
• /review — code review of optimization changes (useful after a complex fix)
• /spec — formalize performance requirements before starting if the target is unclear