Which instruction, memory and skill files each AI coding harness loads, from which paths, with what precedence — measured from official docs, one evidence URL per claim.
The ecosystem converged on a handful of convention files (CLAUDE.md, AGENTS.md, SKILL.md, skill directories) without anyone deciding it. This map is the measurement of that convergence: if you write your project's instructions once, where should they live to be read by the most harnesses?
- matrix.md — harness × convention-path grid.
●= loaded as instruction/context file,◆= discovered as a skill directory. - harnesses/ — per harness: loaded paths with scope (global/project) and documented precedence, skill directories, learned-memory behavior, compatibility notes.
- PRECEDENCE.md — the three conflict models (concatenate / first-match / specificity-overrides) for when two instruction files disagree.
- MEMORY.md — which harnesses write learned memory about you, where it lives, and on-by-default vs opt-in.
- COVERAGE.md — what is included, what is deliberately not, and why.
- SYNTHESIS.md — what the convergence means, and its expiry date.
- Every cell links to the doc page it was read from, fetched during the research run (2026-10-05).
As of this snapshot (27 harnesses): AGENTS.md at the project root is read by 23 of 27 — the de-facto instruction interchange. The four holdouts are instructive rather than rebellious: Aider predates the format entirely; Replit Agent and Tabnine read house files (replit.md, .tabnine/guidelines) instead; and Bolt auto-reads lowercase agents.md while documenting no AGENTS.md at all — the convention spreading as a case-variant. For skills the field splits: .agents/skills is discovered by 19, while .claude/skills is discovered by 12, several of those purely for Claude-Code compatibility. Practical read: put instructions in AGENTS.md; put skills in .agents/skills and mirror into .claude/skills if you care about the compatibility tail. Read SYNTHESIS.md for why that is an observation with an expiry date, not a standard.
This repo answers which file each harness reads. The two adjacent questions — what people actually put in those files (2,303 real ones: testing instructions in 75.9%, security in 14.8%) and whether the files change what agents do (four experiments, one of them a null with a stated power bound) — live in context-file-evidence. Read together they say: the format converged fast, the content did not, and the payoff is smaller and more content-dependent than either side of the debate claims.
python3 scripts/generate.py journals/*.jsonlThe generator sorts the files by the wave order journals/MANIFEST.md declares (later waves win on conflicts), so the glob form is safe — the shell's alphabetical expansion is not the merge order. CI regenerates on every data-path push and fails if any committed page differs from what the journals compile to.
Same evidence contract as harness-atlas: no memory answers, dates travel with claims, unknown over guess.
The map says .agents/skills (19/27) and .claude/skills (12/27) together cover the field. scripts/mirror-skills.sh copies a skill folder into both:
scripts/mirror-skills.sh path/to/my-skill [repo-root]Copies, not symlinks — several harnesses walk skill directories without following links. Re-run after edits.
Orvii — Open, Research, Vision, Innovation & Ideas. The atlas maps capabilities, this maps the files they read.
Part of the Orvii research set: harness-atlas · equivalence-notes · provider-reliability · context-file-evidence · retractions · svg-instruments · bench-notes · ts-lto-research.