Skill Lint

Turn contract

Load and verify the canonical Turn contract before ending every turn; the read-only profile substitutes boxes 1–5 (the lint reports only — nothing was edited, committed, pushed, or created; its one sanctioned output is the git add + commit command block printed for the human, never executed). Skill-specific additions:

✓ `bun scripts/skill-lint.mjs` ran and its report block is pasted verbatim
  (a `PASS` verdict, or every `FINDING` line as printed)
✓ The eight authoring rules were checked against each changed SKILL.md
  (`references/DISCOVERY_AND_LINT.md`) — including the two machine-surface
  parity/ordering checks, the internal-skill discovery-exclusion check and the
  published-vocabulary check — and every violation was reported
✓ Nothing was repaired, added or rewritten — the lint reports only
✓ The git add + commit command block is printed as the ABSOLUTE last output

Missing reference → STOP. About to end the turn with any box unchecked? The turn is NOT done — complete the missing box first.

When to use

After any edit to one or more skills/<name>/SKILL.md files (or to a references/ file beside one) in this repo, before committing. The skill is lint-only: it reports, never repairs, and it bumps nothing — skills carry the unified train version, stamped by scripts/release.mjs apply at cut time (D-40-14), and a skill change ships via a .changeset/ intent file.

Step 0 — Orientation

This skill is specific to the agentic-workflow repository. Before doing anything, confirm you are in that repo (presence of skills/ + CHANGELOG.md). If not, stop and tell the user.

Process

1. Run the content checks

The mechanical half is code:

bun scripts/skill-lint.mjs                    # the whole tree — the sweep
bun scripts/skill-lint.mjs --skill <name>     # only the skill(s) you changed

Paste its block verbatim. The seven checks (hidden comments, frontmatter tags, invisible codepoints, unpinned URLs, credential/exfiltration shapes, backslash paths, time-sensitive claims), the shapes each admits and the smells this repo deliberately does not adopt are owned by skill content checks. Exit 0 is PASS, 1 is FAIL, 2 is a usage error — never re-read the block for a verdict the exit code already gave.

2. Lint, warn, never repair

Check all eight AGENTS.md authoring invariants per change discovery and authoring lint. Report every violation without fixing it. These are structure and registration rules (closing block, phase labels, portability, turn contract, plugin registration and its order, internal-skill metadata); the content checks in step 1 do not cover them, and they do not cover the content checks.

3. Print a summary and the next step

Paste both results (the scanner block, then the authoring-rule findings), then print the next step:

Lint reported. Stage and commit:

git add skills/<name>/SKILL.md .changeset/
git commit -m "<type>(skills): <one-line reason>"

A skill change must ship with a .changeset/*.md intent file declaring the Pi package — scripts/lint-changeset.mjs lints it and the version-parity gate enforces it.

Guardrails

• Never change anything in a SKILL.md — this skill reports; the writer is you, the editor, and the version stamper is scripts/release.mjs apply.
• Never commit or push — this skill only lints and reports.
• Never weaken a check to get a green — a finding is repaired in the prose that caused it, or recorded as a decision in references/CHECKS.md; the patterns live in scripts/skill-lint.mjs and are pinned by its suite.
• If a skill file is not tracked by git (new file not yet added), say so in the report — the change-discovery step cannot see it.

Portability (agents other than Claude Code)

This skill is repo-maintenance for agentic-workflow itself, but the repo may be worked on from any agent:

• No slash-command menu — follow this SKILL.md directly after editing any skill, before committing.
• No per-skill model:/effort: — a mid-tier model is enough; this is classification + lint bookkeeping. The content checks are code, so they give the same answer on any model.
• No bun-only assumption — bun scripts/skill-lint.mjs is the form to print. The fallback is node scripts/skill-lint.mjs: same argv, same block.

Relationship to other skills

• Run after any manual edit to a skill or after execute-phase touches a skill file, and before the commit.
• AGENTS.md section "One version: the unified train" is the policy this skill lints against.
• The version-parity gate (scripts/version-parity-gate.mjs) is the CI half: a skills change must ship with a .changeset/ intent file (or a package bump + companion row).
• Context budgets are not this skill's: scripts/check-skill-context.mjs owns them, and references/CHECKS.md records that as a deliberate skip.
• Renames and other breaking surfaces feed docs/workflow/MIGRATION.md, which init-workspace reads when installing.

Done when

• The scanner's block is pasted and every FINDING line is reported for the user to fix before committing — or the verdict is PASS.
• All eight authoring-rule violations (missing → Next: block, S1/"Step" phase labels, a user-invocable: true skill absent from plugin.json, a non-alphabetical machine surface, a drifted Sev/Impact cell, …) are reported for the user to fix before committing.
• The next step is printed — the git add + git commit command, ready to run.