Skip to main content

CLAUDE.md Evaluation

Does your CLAUDE.md actually help your coding agent?

A controlled before-and-after measurement in pinned Docker containers — same agent, same tasks, honest noise-envelope reporting.

No measurable effect0.0 pp± 20.0 pp · n=5×3 · isolated Docker
egress openn=5×3σ ±0.0 ppGemini (experimental)/gemini-2.5-flash

Measurement pipeline

IngestGenerate CLAUDE.mdDerive tasksBuild pinned imageRun ×N (with/without)ScoreExport

Within noise across 4 evaluations under gemini-2.5-flash — no measurable effect detected so far.

Showing last export — commit 44983090, 13 Jun 2026. Re-run PRIMER and re-export to update.

Evaluations