Creates, updates, or prunes an AGENTS.md for any repository by auditing the codebase, detecting non-discoverable gaps, and drafting minimal high-signal instructions that agents cannot infer from reading the code.
90
94%
Does it follow best practices?
Impact
78%
1.06xAverage score across 3 eval scenarios
Passed
No known issues
{
"context": "Tests whether the agent correctly applies the discoverability filter when generating a fresh AGENTS.md, includes only non-discoverable content, uses the mandated output format, and maintains a required audit trail of exclusions.",
"type": "weighted_checklist",
"checklist": [
{
"name": "Exact heading structure",
"description": "AGENTS.md uses the exact heading '# Agent instructions — <project name>' (no renaming or paraphrasing)",
"max_score": 8
},
{
"name": "Scope statement present",
"description": "AGENTS.md has a one-line or two-sentence scope statement immediately after the main heading explaining what non-discoverable content it covers",
"max_score": 8
},
{
"name": "CI flag documented",
"description": "AGENTS.md documents that tests must be run with the '--no-cache' flag (or equivalent) because CI uses it but the package.json script does not",
"max_score": 10
},
{
"name": "Undocumented env var documented",
"description": "AGENTS.md documents REDIS_URL as a required environment variable that is absent from .env.example",
"max_score": 10
},
{
"name": "Deprecated module status documented",
"description": "AGENTS.md documents the status/constraint of src/legacy-payments/ (e.g., that it is under migration or must not be modified), NOT just that it exists",
"max_score": 10
},
{
"name": "No tech stack content",
"description": "AGENTS.md does NOT mention the programming language, runtime, or framework as a general summary (e.g., no 'This project uses TypeScript' or 'Built with Node.js')",
"max_score": 10
},
{
"name": "No directory overview",
"description": "AGENTS.md does NOT describe what the src/ or other directories contain in general terms (e.g., no 'src/ contains the source code')",
"max_score": 8
},
{
"name": "No linter duplication",
"description": "AGENTS.md does NOT restate ESLint or Prettier rules that are already enforced by .eslintrc or tooling config",
"max_score": 8
},
{
"name": "Package manager mismatch documented",
"description": "AGENTS.md documents which package manager should actually be used, resolving the conflict between the README instruction and the lock file present on disk",
"max_score": 8
},
{
"name": "Audit trail file created",
"description": "An audit-notes.md file exists listing items that were considered but excluded, each with a one-line reason citing the repo file that makes it discoverable",
"max_score": 10
},
{
"name": "Correct section headings used",
"description": "All sections present in AGENTS.md use the prescribed headings verbatim (e.g., '## Non-discoverable commands', '## Landmines', '## Conventions not enforced by tooling') without renaming or paraphrasing",
"max_score": 10
}
]
}