Skillpack Maintenance

Maintain useful task outcomes and actual repository contracts. A skill is not valuable merely because an agent follows it or its files parse.

Establish scope and custody

Read the task, repository instructions, current diff, relevant manifests and existing checks. Preserve unrelated work. Infer clear authorized scope; ask only when a consequential decision is unresolved. Treat a review as read-only unless changes are requested. Existing authorization persists through routine phases.

Inventory relevant published components separately from test fixtures, generated wrappers, and historical reviews. The deliberately flawed plugin under .test-fixtures/flawed-plugin/ is a negative test corpus, not a published pack.

Evaluate incremental value

Identify the concrete failure the component prevents, the project/tool knowledge it supplies, and the artifact or check it enables. Compare against the model's baseline and available native tools. Prefer a short evidence/output contract, selective reference retrieval, or executable validation over generic lectures, mandatory ceremonies, model personas, and duplicate framework introductions.

For substantial behavior changes use representative scenarios with the same model/tools under no skill, current skill, and concise candidate. Score task correctness, verification, user friction, cost/latency, and regressions. Separate model behavior from tool failure. If no comparative run is practical, report that limit; do not claim measured uplift. See testing-skill-quality.md.

Implement within the approved task

1. Trace callers, wrappers, agents, references, packaging, and any parsed output schemas before deleting or changing a public contract.
2. Rewrite the canonical source, remove obsolete duplicate instructions, and update consumers together. Reference facts and examples remain on demand.
3. Keep descriptions discoverable by task/symptom. Use actual platform metadata conventions; tool declarations are restrictions in some runtimes.
4. Validate syntax, local links, catalog/manifest consistency and changed workflows. Run repository checks appropriate to the change, including python scripts/check_marketplace_integrity.py when working in this repository.
5. Update versions, changelogs, and generated wrappers as required by the current repository release policy; do not hand-edit generated outputs blindly.

No external authoring plugin is mandatory. An installed authoring/evaluation skill can help, but its absence does not block a tested change. Do not require uniform SME headings or a particular model family unless a real consumer needs them. Honor the active platform's available tools and permission boundaries.

References and report

• analyzing-pack-domain.md: scope, inventory, utility.
• reviewing-pack-structure.md: packaging and consumers.
• testing-skill-quality.md: outcome-based comparisons.
• implementing-fixes.md: edits, validation, release custody.

Load only the relevant section. Report material changes, exact checks and results, unresolved coverage, and release state. Distinguish source edits, local validation, commit, publication, and actual user acceptance.