quantify-impact
Installation
SKILL.md
Make value obvious without benchmark theater. Every PR creation/update runs this assessment automatically through /commit-push-pr or /stacked-prs; measurement stays proportional.
Flow
- Evidence opportunity scan: benchmark only when a cheap direct metric could clear a predeclared worthwhile delta and normal variance. Tiny copy/style/test work gets a value sentence. No benchmark theater.
- Lock claim before coding: thesis, primary metric, guardrail, scenario, minimum worthwhile delta. Never pick the winner afterward.
- Product lane + Codebase lane: one must improve; the other must not materially regress.
- Product: capability, task success, repro, errors, steps, latency, resources.
- Codebase: maintenance surface, complexity, dependencies, warnings, leaks, bundle, build/test cost, testability.
- Proportional rigor: sentence for obvious value; deterministic repro/count for correctness; controlled paired REFERENCE.md benchmark for runtime; always measure explicit performance claims.
- Base: measure before coding or reconstruct merge-base. Use the same scenario, fixture, config, and machine for base/candidate.
- Compare: only threshold-clearing metrics report raw before/after, absolute/percent delta, method, environment, and noise. Suppress below-threshold/within-variance numbers. Invariant tests/proxies are not performance claims.
- Decide:
- Worthwhile gain:
Value proven. - Ambiguous/negligible explicit performance claim:
Value not proven; no micro-deltas or metric-shopping. Allow one evidence-driven revision, then scrap/close. - No useful metric and no performance claim: normal value summary, no impact artifact.
- Regression: fix, narrow, or stop.
- Worthwhile gain:
Apply the same filter to guardrails; omit negligible movement.