Agents Context Management

Create and polish repo-local context as one system: human README files, agent-facing AGENTS.md files with companion CLAUDE.md symlinks, and existing project-installed skills under .agents/skills.

Workflows

Choose exactly one workflow.

Default to polish when the user asks for updates or invokes the skill without a clear creation request. Default to create only when the user says create, init, initialize, generate, or new for README.md / AGENTS.md context files.

Arguments

• path: Optional repo-relative subtree. Restrict README.md, AGENTS.md, CLAUDE.md, package-root, and project-skill discovery to that subtree.
• skill-name ...: Optional filters for the existing project-installed skills under .agents/skills/<name>/. Applies only to project skill targets during polish.
• --root-only: Restrict README.md / AGENTS.md creation or polishing to the repository root. Does not select nested .agents/skills targets.
• --dry-run: Preview planned writes and diffs without changing files.
• --preserve: For polish, keep user-authored prose and structure where still accurate; fix only drift and obvious noise.
• --minimal: Use the smallest useful README.md / AGENTS.md output.
• --thorough / --full: Perform deeper analysis and include more detail when it adds durable context.
• --force: For create, overwrite existing README.md or AGENTS.md targets without prompting. Never applies to skills.

Surface unrecognized flags in the final report and continue with defaults when safe.

Scope

Stay inside the current git repository. Never scan or write global installs under ~/.agents, ~/.claude, or any .claude/skills directory.

This skill may write:

• Existing README.md files during polish.
• Missing README.md files during create.
• Existing AGENTS.md files during polish.
• Missing AGENTS.md files during create.
• Sibling CLAUDE.md symlinks that mirror an AGENTS.md in the same directory.
• Existing repo-local .agents/skills/<name>/SKILL.md files, plus their existing references/, scripts/, assets/, and examples/ files when a minimal factual fix requires it.

This skill must not:

• Create, install, delete, or rename skills. For new skills, refer to skills/create-skill.
• Treat catalog directories such as skills/<name>/ as project-installed skills.
• Process CLAUDE.md as a source file unless it is a symlink to sibling AGENTS.md.
• Rewrite regular CLAUDE.md files.
• Edit CONTRIBUTING.md; only recommend merging it into sibling AGENTS.md.
• Auto-commit.

Guard Rails

Run these checks before discovery or writes:

cwd="$(pwd -P)"
case "$cwd" in
  /) printf 'abort: refusing to run at the filesystem root\n' >&2; exit 1 ;;
  "$HOME/.agents"|"$HOME/.agents/"*|"$HOME/.claude"|"$HOME/.claude/"*)
    printf 'abort: refusing to run under ~/.agents or ~/.claude\n' >&2; exit 1 ;;
esac
repo_root="$(git rev-parse --show-toplevel 2>/dev/null)" || {
  printf 'abort: not inside a git repository\n' >&2; exit 1; }
case "$repo_root" in
  /|"$HOME") printf 'abort: unsupported repo root: %s\n' "$repo_root" >&2; exit 1 ;;
  "$HOME/.agents"|"$HOME/.agents/"*|"$HOME/.claude"|"$HOME/.claude/"*)
    printf 'abort: repo root is under ~/.agents or ~/.claude\n' >&2; exit 1 ;;
esac

Snapshot git status --short before broad edits. Ignore unrelated pre-existing changes, but do not overwrite them blindly.

Discovery

Use git-aware discovery for documentation targets:

git -C "$repo_root" ls-files --cached --others --exclude-standard -- \
  '**/README.md' 'README.md' '**/AGENTS.md' 'AGENTS.md'

For creation, package roots are the repo root plus directories containing a manifest:

git -C "$repo_root" ls-files --cached --others --exclude-standard -- \
  '**/package.json' 'package.json' '**/Cargo.toml' 'Cargo.toml' \
  '**/pyproject.toml' 'pyproject.toml' '**/setup.py' 'setup.py' \
  '**/go.mod' 'go.mod' '**/foundry.toml' 'foundry.toml' \
  '**/Gemfile' 'Gemfile' '**/composer.json' 'composer.json'

Discover existing project-installed skills deliberately, including ignored .agents/ directories:

fd --glob --full-path --hidden --no-ignore --follow --type f \
   --exclude .git --exclude .claude --exclude node_modules --exclude vendor \
   --exclude dist --exclude build --exclude out --exclude target \
   --exclude .next --exclude .venv --exclude coverage \
   '**/.agents/skills/*/SKILL.md' "$repo_root" \
| while IFS= read -r p; do
    dir="$(cd "${p%/SKILL.md}" 2>/dev/null && pwd -P)" || continue
    case "$dir" in
      "$HOME/.agents/"*|"$HOME/.claude/"*) continue ;;
      "$repo_root"/*) printf '%s/SKILL.md\n' "$dir" ;;
    esac
  done \
| awk '!seen[$0]++'

Skip VCS, dependency, and build output directories: .git, node_modules, vendor, .venv, target, dist, build, out, .next, and coverage. Skip hidden dot-directories unless they contain a manifest or are the explicit .agents/skills target set.

Core Rules

• Treat README.md as human-facing. Keep description, badges, links, references, acknowledgments, license, and a pointer to AGENTS.md. Keep CLI commands only for short operator-run setup guides such as dotfiles, infra, homelab, or explicit user requests.
• Treat AGENTS.md as agent-facing. Keep commands, non-obvious architecture, style rules, safety constraints, generated-file ownership, workflow traps, and contribution guidance.
• Treat CLAUDE.md as a compatibility symlink. Create or refresh it with (cd "$dir" && ln -sf AGENTS.md CLAUDE.md) only when it is missing or already a symlink.
• Treat project-installed SKILL.md files as existing workflows to verify, not content to redesign. Fix factual drift with the smallest edit that restores truth.
• Prefer deletion over summarizing noise. Remove content that can be cheaply inferred from manifests, filenames, or standard tool behavior.
• Keep parent AGENTS.md files for global defaults and nested AGENTS.md files for local deltas.
• During polish, recommend missing nested AGENTS.md files only when a subtree has distinct commands, safety rules, generated files, ownership boundaries, or data-handling constraints. Create them only through the create workflow.

Verification

After writes:

1. Run the host repo's Markdown formatter/checker. Prefer just mdformat-write, then just mdformat-check when present.
2. If any SKILL.md frontmatter or agents/openai.yaml changed in a skills catalog, run the host repo's invocation metadata check. Prefer just skill-invocation-check when present.
3. Verify changed CLAUDE.md symlinks resolve to sibling AGENTS.md.
4. For --dry-run, skip writes and verification that depends on changed files; report the exact commands that would run.

Reporting

Report in this order:

1. Scope: target counts and relative paths for README.md, AGENTS.md, CLAUDE.md symlinks, and existing project skills.
2. Changes: grouped by relative directory or .agents/skills location.
3. Advisories: CONTRIBUTING.md merge notes, obsolete skills, skipped regular CLAUDE.md files, unrecognized flags.
4. Verification: exact commands run and outcomes.
5. Residual Risks: None. or one terse line per risk.

Use ✓ for updated/created/verified, ⊘ for skipped, ⚠ for advisories, and ✗ for failures.

References

• Read references/brain-polish.md for the unified update/polish workflow.
• Read references/create-docs.md for README.md and AGENTS.md creation.