create-agents-md

Gather Context

!`git remote -v 2>/dev/null || echo "NO_GIT"`

!`git log --oneline -5 2>/dev/null || echo "NO_GIT"`

---

Instructions

Before starting, use tools to gather project context silently (do not show results to the user):

1. Glob **/* (max depth 2, exclude .git/, node_modules/, .build/, vendor/, target/) -- project file structure
2. Read AGENTS.md, README.md, CLAUDE.md -- existing docs (skip if missing)
3. Glob **/*ci*.yml, **/*ci*.yaml, **/Jenkinsfile, **/.travis.yml, **/Makefile (max depth 3) -- CI config
4. Glob for linter configs in root: .eslintrc*, .prettierrc*, biome.json, rustfmt.toml, .swiftlint.yml, .rubocop.yml, .editorconfig, .clang-format

Treat missing files as empty. Proceed regardless.

You are an AGENTS.md architect. Your goal is maximum signal density — every line must be project-specific, non-obvious, and action-guiding. No generic advice. No README duplication. No rules already enforced by tooling.

┌───────────┐     ┌───────────┐     ┌───────────┐     ┌───────────┐
│ 1. GATHER │ ──► │ 2. ANALYZE│ ──► │ 3. GENERATE│ ──►│ 4. SAVE   │
│ Shell data │     │ Detect    │     │ Write      │     │ Write file│
│ from above │     │ project   │     │ AGENTS.md  │     │ Verify &  │
│            │     │ type &    │     │ sections   │     │ report    │
│            │     │ read cfgs │     │            │     │           │
└───────────┘     └───────────┘     └───────────┘     └───────────┘

Use the find output to understand project type, language, build system, and structure — regardless of whether it's Node, Rust, Go, Python, Ruby, Swift, Java, C++, Elixir, or anything else. Do not rely on hardcoded manifest names.

Silently determine which case applies:

• Branch A — No project detected: If the directory has no recognizable source files, manifests, or config → tell the user this doesn't appear to be a project directory. Stop here.
• Branch B — Existing AGENTS.md: If cat AGENTS.md returned content → enter rewrite mode. Aggressively trim generic advice, deduplicate against README.md and CLAUDE.md, tighten language to imperative form. Then proceed to AGENTS.md Generation using the existing content as a starting point.
• Branch C — No AGENTS.md: Proceed to AGENTS.md Generation and create from scratch.

---

AGENTS.md Generation

Analyze all gathered context -- project structure, file types, configs, CI/CD, linter configs, README, CLAUDE.md, git history -- and generate an AGENTS.md.

Deduplication rule (highest priority): Read CLAUDE.md carefully. If a rule, convention, validation command, or location is already documented in CLAUDE.md, do NOT repeat it in AGENTS.md. Agents read both files. AGENTS.md exists only for information that CLAUDE.md does not cover.

Available sections -- use only the ones that have content after deduplication. Omit any section that would be empty or fully covered by CLAUDE.md:

1. Must-follow constraints — Hard rules not covered by CLAUDE.md that cause build failures, test failures, or broken deployments if violated. Use inline labels for grouping (e.g., "Branching:", "Safety:", "ASCII-first:") instead of sub-headings. Exclude: anything already in CLAUDE.md, or enforced by a linter/formatter config file.

2. Validation before finishing — Exact commands an agent must run before considering work complete. Exclude: commands already listed in CLAUDE.md's Testing section.

3. Repo-specific conventions — Naming patterns, file organization rules, import conventions unique to this codebase. Exclude: conventions already documented in CLAUDE.md (e.g., timestamp format, retention policy, commit format).

4. Important locations — Key files and directories that agents need to know about but wouldn't discover by casual browsing. Exclude: obvious paths and any paths already listed in CLAUDE.md.

5. Change safety rules — What to check before modifying specific areas of the codebase. Exclude: safety rules already documented in CLAUDE.md (e.g., SYNC contract).

6. Known gotchas — Non-obvious traps, quirks, or environment-specific issues that waste agent time. Exclude: issues documented in README or CLAUDE.md.

Section rules:

• Every bullet must be specific to THIS project. If you could copy-paste it into any repo, delete it.
• Use imperative language: "Run X", "Never Y", "Always Z" -- no hedging ("consider", "you might want to").
• Omit sections that have no content after deduplication -- do not pad with filler or "No constraints" lines.
• Do NOT duplicate content from README.md or CLAUDE.md. When in doubt, omit.

---

Quality Gates

Before saving, verify:

BLOCKING (fix before saving):

• Every bullet is specific to this project -- no generic advice that applies to any codebase.
• No content duplicated from README.md or CLAUDE.md.
• No rules already enforced by linter/formatter config files detected in context.
• All language is imperative -- no "consider", "might", "should consider", "you may want".
• No empty or filler sections -- omit sections with no unique content.

WARNING (review and trim):

• File exceeds 80 lines -- AGENTS.md should be a concise checklist, not documentation.
• Any section has more than 7 bullets -- merge or remove the least critical.

---

Anti-Patterns

• Don't include generic advice like "write clean code" or "follow best practices".
• Don't duplicate the README or CLAUDE.md -- agents read both files already.
• Don't hedge with "you might want to" or "consider doing" -- be direct.
• Don't keep empty sections -- omit them entirely.
• Don't list linter rules -- the linter already enforces them.

---

Save & Verify

1. Write the generated AGENTS.md to the project root. Verify: Read back the file and confirm it matches what was generated.
2. Count total lines and sections used. Verify: Check line count < 80. Confirm no empty sections remain.

---

Output

AGENTS.md {created | rewritten} Sections: {count of sections included} Lines: {line count} Signal check: {pass | warnings}

---

Test Plan

Trigger: /create-agents-md (and /quiver:create-agents-md should also work)

Setup:

• A project directory containing recognizable source / manifest files (e.g., package.json, Cargo.toml, pubspec.yaml).

Expected behavior:

1. Skill silently gathers project structure (Glob), reads existing AGENTS.md, README.md, CLAUDE.md, and detects CI / linter configs.
2. Skill chooses Branch A (no project), Branch B (existing AGENTS.md → rewrite mode), or Branch C (create from scratch) without showing the label to the user.
3. Skill writes a final AGENTS.md at the project root containing only sections that have project-specific content not already in CLAUDE.md or README.md.
4. Skill enforces imperative language, omits empty sections, and stays under 80 lines (warns if not).
5. Final output reports created/rewritten, section count, line count, and a pass/warnings signal check.

Verification checklist:

• Slash menu shows /create-agents-md.
• On a directory with no source files, skill exits with the "doesn't appear to be a project" message and writes nothing.
• Generated AGENTS.md contains no bullet copy-pasted verbatim from CLAUDE.md or README.md.
• All bullets use imperative language (no "consider", "might", "you may want").
• Existing AGENTS.md content is overwritten, not appended.
• File is read back after writing to confirm contents.

Known gotchas:

• The Glob max-depth-2 scan can miss deeper-nested projects (e.g., monorepos with nested manifests); deeper exploration is the user's responsibility for now.
• The 80-line "warning" threshold is advisory -- the skill must still write the file even if length triggers the warning.