AGENTS.md Skill
Empirically-grounded guidance for writing agent instruction files (AGENTS.md, CLAUDE.md,
etc.). Based on AuggieBench eval results (Augment Code, Apr 2026).
Core finding: The same file can improve one task by 25% and tank another by 30%. What you leave out matters as much as what you put in.
Reference files
refs/patterns.md— The six effective patterns with examples and data. Read when writing or improving an AGENTS.md.refs/failure-modes.md— Four failure modes with diagnosis and fixes. Read when auditing a broken or underperforming AGENTS.md.
Workflow: Writing an AGENTS.md
Identify the module scope. Count core files. Check for surrounding doc sprawl (specs, architecture docs nearby). Sprawl is an environment problem — fix it before fixing the entry point.
Pick the right size target.
Module size AGENTS.md target Reference files Small (<30 files) 50–80 lines 0–1 Mid (~100 files) 100–150 lines 2–5 Large (200+) 100–150 lines + refs 5–10, scoped Do not scale length with module size. Scale the reference structure instead.
Identify which problems exist. Use the decision table below to pick patterns.
Write the entry point. Apply patterns from
refs/patterns.md. Keep architecture prose to 2–3 sentences max. Describe boundaries, not internals.Push depth into reference files. Each ref needs a clear scope statement in
AGENTS.mdso the agent knows exactly when to open it. Max 10–15 refs total.Audit for failure modes. Check against
refs/failure-modes.mdbefore shipping.
Workflow: Auditing an existing AGENTS.md
- Check the environment first. Does the module have large surrounding doc sprawl? If yes — that's likely the primary problem, not the file itself.
- Measure line count. Over 150 lines → start splitting into reference files.
- Count "don'ts" without "dos". More than 5 → warning overload risk.
- Look for architecture prose. Descriptions of why things are structured a certain way → overexploration trigger. Cut or move to refs.
- Check for net-new patterns. If the task requires a pattern not in the codebase, AGENTS.md will steer wrong. Needs a spec, not docs.
Decision table: which pattern to apply
| Symptom / Goal | Pattern | Details |
|---|---|---|
| Agent reinvents existing abstractions | Real codebase examples | refs/patterns.md §4 |
| Agent picks wrong library or approach | Decision table | refs/patterns.md §3 |
| Agent ships features with missing wiring | Procedural workflow | refs/patterns.md §2 |
| Agent ignores codebase conventions | "Don't" + "Do" pairs | refs/patterns.md §5 |
| Agent over-explores before writing code | Progressive disclosure; trim arch prose | refs/failure-modes.md §1 |
| Removing AGENTS.md barely changes behavior | Audit surrounding doc sprawl | refs/failure-modes.md §3 |
| Agent builds the wrong architecture | Spec-driven development first | refs/failure-modes.md §4 |
Document discovery rates
| Location | Read rate |
|---|---|
AGENTS.md (full hierarchy) |
100% |
Files referenced from AGENTS.md |
~90% |
README.md in working directory |
~80% |
| Nested READMEs in subdirectories | ~40% |
Orphan docs in _docs/ (unreferenced) |
<10% |
If it needs to be seen, it lives in AGENTS.md or is directly referenced from it.