Github curator

github-curator

The meta-skill. This skill curates a repository of Agent Skills, including the repository it lives in. It is the first and canonical example of its own output.

Why this skill exists

Skill repositories rot. Frontmatter drifts from the spec, AGENTS.md falls out of sync with CLAUDE.md, version numbers get bumped without changelog entries, packets ship without IP scrubs, and the four-gate publish-readiness pattern becomes ceremonial rather than enforced. The github-curator skill prevents drift by making the repo verify itself on every change and by treating the repo as its own first audited artifact.

The Anthropic Agent Skills specification at https://agentskills.io/specification (released 2025-12-18) is the source of truth for skill structure. The AGENTS.md standard at https://agents.md is the source of truth for repo-root agent instructions. This skill enforces both, and rejects any composition that deviates from them.

When to use

Load this skill when the agent is asked to do any of the following on a Skills-pattern repo:

1. Add a new skill folder.
2. Edit an existing skill's SKILL.md or reference docs.
3. Cut a SemVer release.
4. Produce a packet zip for distribution.
5. Audit the repo against the Agent Skills spec or the AGENTS.md standard.
6. Onboard a new contributor (humans or agents) to the curator pattern.
7. Resolve drift between AGENTS.md and CLAUDE.md.
8. Resolve drift between registry.yaml and the actual skills/ tree.

Do not load this skill for unrelated repository tasks.

The loop

This skill runs a four-step loop. Every step writes to the file system. None of the steps mutate state without producing a verification record.

READ -> WORK -> WRITE -> STOP

The loop is taken verbatim from system_directive_protocol Section 4 (Episodic Runtime Loop). It is the same loop the ingest skill uses on external bundles. The point of using the same loop in both places is that this repo is auditable by ingest itself, the repo IS an external bundle to its own ingest skill, and the recursion terminates correctly.

READ (active sensing)

1. Read AGENTS.md at repo root.
2. Read registry.yaml at repo root.
3. Read each skills/<n>/SKILL.md. Confirm frontmatter is valid:
   • name field is present, lowercase-hyphenated, ≤64 chars, equals the parent directory name.
   • description field is present, between 1 and 1024 characters.
   • license and compatibility are optional but recommended.
4. Confirm CLAUDE.md content is the 3-line pointer to AGENTS.md (no divergent instructions).
5. Confirm CHANGELOG.md has an [Unreleased] section.

If any READ step fails, STOP and surface the failure. Do not proceed to WORK.

WORK (reasoning)

The agent may now plan the requested change. Planning produces a write-list: every file path that will be modified, created, or deleted in WRITE.

The write-list must be reviewable in plain text. No silent file modifications.

WRITE (artifact generation)

1. Apply the write-list.
2. Update CHANGELOG.md [Unreleased] with one entry per modified skill. Use Keep a Changelog categories: Added, Changed, Deprecated, Removed, Fixed, Security.
3. If a new skill was added, append a row to registry.yaml and a row to skills.md if the parent repo also maintains a tree index.
4. Run the verification matrix (next step).

STOP (termination with verification)

Run skills/github-curator/scripts/verify-self.sh. The matrix output must be PASS before the change is considered done. Any FAIL halts the loop. The loop does not retry automatically; the agent surfaces the failure to the human.

The verification matrix

The matrix is a boolean-atomic check list. Each row returns TRUE or FALSE. Each row has one importance tag (Essential / Important / Optional / Pitfall) following the Rubrics-as-Rewards pattern (arXiv 2507.17746). The full matrix lives in references/self-verification-matrix.md. Load it on demand, not at skill-activation time.

Summary of the rows (full text in references):

#	Tag	Check
1	Essential	AGENTS.md exists at repo root.
2	Essential	Each skills/<n>/SKILL.md frontmatter validates.
3	Essential	Skill name equals parent directory name.
4	Essential	description length is between 1 and 1024 chars.
5	Essential	LICENSE exists and is a recognized open license.
6	Essential	No em dashes (U+2014) and no en dashes (U+2013).
7	Essential	No tier-1 IP scrub hits (see publish-readiness-rubric).
8	Important	CLAUDE.md is the 3-line pointer to AGENTS.md.
9	Important	CHANGELOG.md [Unreleased] reflects the diff.
10	Important	registry.yaml rows match the skills/ tree.
11	Important	Conventional Commits message proposed for the change.
12	Optional	Compatibility field lists at least 3 host environments.
13	Optional	Reference docs follow progressive-disclosure conventions.
14	Pitfall	No skill loads more than one body at activation time.
15	Pitfall	No reference doc is loaded eagerly in SKILL.md.

Any row tagged Essential that returns FALSE blocks the WRITE. Important rows that return FALSE produce a warning; the agent surfaces them but proceeds at the human's explicit confirmation. Optional rows surface as flags. Pitfall rows that return TRUE are bugs that must be repaired before the change lands.

Self-application: this skill on this repo

This skill was the first artifact this repo produced. The verification matrix has been run against this repo. The result on bootstrap was 15 PASS / 0 FAIL. The output lives in scripts/verify-self.sh and reproduces that result on every commit.

The dogfood is complete: this SKILL.md is the first row in the registry.yaml, the AGENTS.md describes the build/test/release commands that scripts/ implements, the LICENSE matches the frontmatter license: field, and CLAUDE.md is the 3-line pointer described in AGENTS.md. The repo passes its own gate.

Composition with other skills

The github-curator skill calls three siblings:

• ingest runs at WORK step when an external bundle is being absorbed into the repo.
• publish-readiness-rubric runs at STOP step as the boolean-atomic verification matrix.
• session-dispatch runs when the change is large enough to warrant a fresh chat session (the human spawns a new workstream).

A typical multi-skill turn loads github-curator plus exactly one of the three siblings. Loading more than one sibling body at once is a Pitfall row in the matrix.

Versioning rules

The repo follows SemVer 2.0.0 (https://semver.org):

• PATCH: typo fixes in skill bodies, link fixes, reference doc clarifications.
• MINOR: new skill added, new reference doc, new packet, new compatibility entry.
• MAJOR: removed skill, moved file path that downstream packets reference, breaking change to a SKILL.md frontmatter (renamed name, removed required field), or a change in the four-gate semantics that downstream consumers depend on.

A release is cut by make release VERSION=vX.Y.Z, which produces a git annotated tag, a packets/github-curator-vX.Y.Z.zip, and a checksums.txt containing sha256 lines per asset. The zip filename matches the directory name exactly, per the vercel-labs/agent-skills convention.

Distribution

The repo is "Use this template" enabled on GitHub. Forks get a clean history. Each release attaches the zip to a GitHub Release at https://github.com/austinbgreen/github-curator/releases. Drop the zip's contents into NotebookLM, a Claude project, a custom Gem, or a custom GPT, and the skills behave identically: the SKILL.md format is host-neutral, only name plus description load on activation, and the references load on demand.

When this skill should fail closed

The Zero Assumption Mandate from system_directive_protocol Section 3 is hard. If any of the following is true, this skill stops and asks the human:

1. A SKILL.md name field does not match the parent directory.
2. A description is missing or longer than 1024 characters.
3. A skill is referenced in registry.yaml but its directory does not exist.
4. A skill directory exists but is missing from registry.yaml.
5. CLAUDE.md contains divergent instructions from AGENTS.md.
6. The IP scrub gate (tier-1) hits on any file in the repo.
7. The repo contains an em dash or en dash (rule 3 of pre_submit_gate).

The skill does not attempt to repair these silently. Each is a STOP condition with a numbered question to the human.

Reference docs

Loaded on demand, not at activation:

• references/self-verification-matrix.md the full 15-row matrix.
• references/agent-skills-spec-conformance.md the agentskills.io spec fields this repo enforces, with examples.
• references/agents-md-conformance.md the AGENTS.md standard fields this repo includes, with the discovery-order rules.
• references/release-procedure.md the SemVer + GitHub Releases procedure, including the make release script breakdown.
• references/upstream-lineage.md the list of repos and standards this bundle composes (vercel-labs/agent-skills, anthropics/skills, anthropics/claude-cookbooks, openai/openai-cookbook, agents.md, agentskills.io, semver.org, keepachangelog.com).

Provenance

This SKILL.md was authored 2026-04-26 by 0SxD as the first publishable composition of the system_directive_protocol governance set onto the open Agent Skills standard.

END_GITHUB_CURATOR_SKILL