optimize-accuracy-first

Optimize Accuracy First

Optimize the work required to produce a correct, useful result. Do not optimize one local timer while making callers search again, consume worse evidence, or lose a correctness guarantee.

This skill is self-contained. Use repository source, version control, the project's normal build and test commands, checked-in benchmarks, and ordinary OS measurements. Do not require an optional MCP server, hosted research service, or external workflow product. If an optional facility is unavailable, continue with the repository-owned workflow.

Read references/measurement-matrix.md before designing or running an experiment.

Define the actual outcome

Write one falsifiable optimization claim before changing code:

For workload W, change C should reduce owner O and improve metric M, while preserving invariants I.

Name the caller-visible outcome. “Reduce SQL statements” is not enough if the same rows still dominate. “Use less CPU” is not enough if retrieval misses cause another request. Prefer claims such as:

• reduce verified regex chunks without changing matches or bounded failures;
• improve relevant-source coverage within the same exact token budget;
• remove repeated content scans while preserving excerpts, ordering, and scores;
• reduce cold publication writes without weakening atomic generations;
• reduce deep-read work without changing live-file staleness detection.

Classify the request before acting:

• For diagnose, locate and explain the owner; do not implement.
• For research or compare, produce a decision with evidence; do not mutate.
• For change or optimize, implement the narrowest supported candidate and validate it.

Establish non-negotiable invariants

Record the applicable invariants before measuring. Include more than happy-path result equality:

• result membership, coordinates, ordering, scores, and occurrence counts;
• exact token accounting, truncation, continuation, and saturation semantics;
• case, Unicode, regex, parser, and tokenizer semantics;
• error categories, limits, cancellation, and LimitExceeded behavior;
• deterministic output and stable tie-breaking;
• live-content freshness, expected-hash behavior, and snapshot consistency;
• atomic publication, generation invalidation, and bounded memory;
• protocol shape and downstream caller expectations.