agents-md-fixer

AGENTS.md fixer

Goal

Make AGENTS.md a tiny pointer file. It should mostly be bare path pointers into docs/*.md, not prose. The longer AGENTS.md grows, the worse AI coding agents perform: they skim, prioritize poorly, and miss the load-bearing rules. This skill audits AGENTS.md and returns an aggressively tightened version, with the manager (the invoking agent) exercising judgment about what is important for this repo versus frivolous.

AGENTS.md is written for AI agents, not humans. Do NOT use Markdown links ([text](path)); reference docs by bare path (docs/PYTHON_STYLE.md) so the file stays plain for the agents that read it. Clickability on GitHub is a human concern and is not a goal here.

Size guidance (line count is a rough proxy, not a contract; "line" means a non-blank text line in the rendered Markdown):

• Prefer about 15 lines or fewer.
• Hard cap at about 50 lines. If the file is longer, something belongs in docs/*.md instead.
• The shorter the better. A 10-line AGENTS.md that links to the right docs is better than a 40-line one that restates them.

If a rule has a canonical home in docs/*.md, point to it by path, do not restate. The docs/REPO_STYLE.md "AGENTS.md files" section describes the philosophy.

Required content (keep)

Keep only items that are repo-specific and load-bearing:

• Bare path pointers to canonical style docs that exist in this repo, for example docs/REPO_STYLE.md, docs/PYTHON_STYLE.md, docs/MARKDOWN_STYLE.md, docs/PYTEST_STYLE.md, and docs/CHANGELOG.md.
• Repo-specific runtime quirks: interpreter bootstrap, env activation, container entry, non-default install paths.
• Non-obvious workflow constraints actually enforced in this repo (example: "humans run git commit", branch policy, hook behavior).
• Test invocation cheat-line specific to this repo.
• Any user-injected directive marked as overriding defaults (preserve it verbatim).

Frivolous content (cut or move)

Flag and remove these patterns:

• Long philosophy or "why we believe X" paragraphs. Move the canonical version to docs/REPO_STYLE.md, or delete if already covered there.
• Restated style rules already in docs/PYTHON_STYLE.md, docs/MARKDOWN_STYLE.md, or docs/REPO_STYLE.md.
• Aspirational guidance with no enforcement and no concrete action.
• Generic AI etiquette unrelated to this repo (tone, manners, hedging advice).
• Commit-message and changelog rules already in docs/CHANGELOG.md or docs/REPO_STYLE.md.
• "Future considerations" sections with no owner and no date.

Manager judgment rubric

For each line or section, answer in order. Stop at the first "yes" that demands action.

1. Is this repo-specific (would it differ in a sibling repo)? If no, cut or replace with a link to the canonical doc.
2. Is the same rule already written in a docs/*.md file? If yes, replace with a link to that file.
3. If this line were removed today, would an agent's behavior change on a real task? If no, cut.
4. Is the canonical home a sibling doc that already exists? If yes, move the content there and link from AGENTS.md.
5. Otherwise, keep, and trim wording to a single short bullet.

Workflow

1. Read the current AGENTS.md and the docs/ index.
2. Get a rough size: wc -l AGENTS.md. Treat this as a smell test, not a contract. If the file is already a tight pointer file (roughly under 15 non-blank lines) and every section already passes the rubric, report "no action needed" and stop. Do not invent churn.
3. For each section, apply the rubric. Categorize each as one of: keep, link-only, move-to-docs, or delete.
4. For move-to-docs items, identify the right docs/*.md target using the recommended common docs list in docs/REPO_STYLE.md. If the target doc does not exist yet, do not create it here. Note it as a follow-up for docset-updater, setup-install-usage-docs, or arch-docs.
5. Rewrite AGENTS.md minimally: short headings, sentence case, bullet lists with -, and bare doc paths (no Markdown links). AGENTS.md is agent-facing; docs/MARKDOWN_STYLE.md link rules apply to human-facing docs, not here.
6. Preserve any user instruction explicitly marked as overriding defaults.

Markdown house rules

• ASCII only; escape symbols like α if needed.
• Sentence case headings, short headings (3-6 words).
• Bullets use -, one idea per bullet.
• Reference docs by bare path, NOT Markdown links. Good: docs/REPO_STYLE.md. Bad: [docs/REPO_STYLE.md](docs/REPO_STYLE.md). AGENTS.md is read by AI agents, not browsed by humans, so clickability is not a goal; a plain path keeps the file lean and unambiguous.
• A bare path may be wrapped in backticks for readability; do not wrap it in link syntax.

Quality bar

• Final AGENTS.md is small: prefer about 15 non-blank lines, hard cap about 50. Smaller is always better. If you cannot get under 50, more content belongs in docs/*.md.
• Most of the file is bare path pointers into docs/*.md, not prose.
• No Markdown links; docs are referenced by bare path.
• No duplication of canonical docs/*.md content.
• Every external concept is referenced by path, not restated.
• No deletion of genuinely repo-specific operational rules.
• No new sections with content that has no enforcement or no concrete action.

Inputs to request

• Current AGENTS.md content.
• The list of files under docs/ (so the skill can choose link targets that actually exist).
• Any standing user override that must be preserved verbatim.

Output

• A proposed AGENTS.md patch with minimal edits, ready for review.
• A change log with three short lists:
  • moved: each chunk moved into docs/*.md, with the destination file.
  • linked: each chunk replaced by a bare path pointer, with the target.
  • deleted: each chunk deleted, with one sentence of justification.
• A short follow-up list of docs/*.md stubs that should exist, naming the skill that owns each one (docset-updater, setup-install-usage-docs, arch-docs, or readme-docs).

References

• docs/REPO_STYLE.md: "AGENTS.md files" section sets the 100-150 line target and the "concise and operational" rule.
• docs/MARKDOWN_STYLE.md: heading and link conventions used in the rewrite.
• skills/readme-docs/SKILL.md: sibling single-artifact standardizer with the same shape.

Delegated execution

Under delegate-manager-to-subagents, this skill is assigned to a fresh subagent with one bounded task, the relevant repo rules, and one verification step. Do not continue the same subagent across unrelated follow-up work; dispatch a new subagent for each atomic task. See docs/REPO_STYLE.md.