Skill Authoring

Skill Authoring

Inherit the default workflow

Read and apply the installed $skill-creator skill completely before changing a repo-local skill. Apply its naming, initialization, frontmatter, imperative writing, progressive disclosure, resource selection, UI metadata, validation, and forward-testing guidance.

Do not initialize an existing skill again. For a new skill, use the installed initializer once, then replace every placeholder and remove every unused example or resource directory. Keep repository rules additive; do not weaken the default skill requirements.

Author concise engineering contracts

Keep each SKILL.md focused on durable, non-obvious knowledge another agent needs to implement or verify behavior:

• responsibility and exclusions;
• inputs, outputs, state, and lifecycle;
• invariants, ordering, bounds, and decision rules;
• errors, cancellation, partial evidence, cleanup, and recovery;
• security and compatibility boundaries;
• language-neutral acceptance scenarios.

Use imperative or infinitive wording. Keep source-symbol inventories and session transcripts out of the contract. Prefer observable behavior and package ownership; mention an algorithm only when compatibility or correctness depends on it.

Keep detailed variants in a directly linked references file only when they would bloat the body. Keep references one level from SKILL.md, explain exactly when to read each one, and do not duplicate its content in the body. Add scripts only for repeated deterministic work and test every added script. Add assets only when they are consumed in repository output.

Maintain skill metadata

Use a sibling directory named exactly .codex/skills/<skill-name>. Use lowercase letters, digits, and hyphens and keep the name under 64 characters.

Put only name and description in SKILL.md frontmatter. Make the description state both the capability and concrete trigger conditions because it controls discovery before the body loads.

Keep agents/openai.yaml synchronized with the finished skill:

• derive a human-facing display name from the skill;
• keep the short description concise and aligned with the responsibility;
• make the default prompt explicitly invoke $<skill-name> for a representative task;
• include no optional interface fields unless they are deliberately supplied.

Regenerate metadata with the installed $skill-creator generator when it is stale. Do not hand-maintain divergent descriptions.

Build forward-only actionable routing

Store every discoverable skill as a direct sibling under .codex/skills. Express hierarchy only with an actionable Markdown sentence that:

1. starts with the imperative Use;
2. links the exact skill name to its relative SKILL.md;
3. contains to followed by a concrete task or trigger;
4. appears in the narrowest router able to make that decision.

Treat only that complete sentence form as hierarchy. A supporting prose link is not a route. Do not add parent metadata, ancestry sections, routing-only backlinks, or a leaf routing section with no delegates.

Keep this canonical topology:

• route only implementation-architecture, skill-authoring, and source-code-management from AGENTS.md;
• route coding-directives, implementation-test-pipeline, implementation-coverage, implementation-reporting, and scripts-and-automation from implementation-architecture;
• keep those specialized skills as leaves unless a future cohesive domain justifies a forward child.

Ensure every route points forward, every target exists, every action is selective, every sibling is reachable from AGENTS.md, and no cycle exists. Put shared invariants in the router and specialized behavior in its leaf rather than copying the same contract into both.

Revision workflow

1. Read AGENTS.md, inventory every sibling skill and agents/openai.yaml, and follow all current actionable routes.
2. Read the source, tests, raw logs, or user correction that establishes the durable behavior. Distinguish required behavior from implementation status.
3. Select the narrowest existing owner. Add a new skill only when loading the concern independently improves cohesion or context use.
4. Edit the contract in imperative form. Mark unimplemented or partial behavior honestly instead of describing intended automation as present.
5. Add or change one forward actionable route at the narrowest dispatch point.
6. Refresh UI metadata when responsibility or triggering changed.
7. Remove all initializer instructions, unfinished markers, example placeholders, unused resources, duplicate guidance, and routing-only backlinks.
8. Run make skills-check. Treat this dependency-free repository validator as the CI authority for frontmatter, UI metadata, placeholders, exact route targets, canonical topology, reachability, and cycle detection. When the installed quick_validate.py and its YAML dependency are available, run it against every changed skill as a supplementary compatibility check.
9. Review the validator's route inventory starting at AGENTS.md; resolve any missing target, ambiguous action, topology mismatch, cycle, or unreachable sibling instead of weakening the check.
10. Report the repository-relative routing chain and every validator result.

Validation expectations

Treat frontmatter validation as the minimum, not the complete audit. Also confirm:

• directory and frontmatter names match;
• frontmatter contains only the two allowed fields;
• descriptions contain concrete triggers;
• bodies are imperative, concise, and free of placeholders;
• every agents/openai.yaml matches its skill;
• all sibling skills are reachable through the canonical forward graph;
• specialized implementation contracts retain raw-evidence, status, incompleteness, security, compatibility, determinism, and verification requirements relevant to their owner.

Keep scripts/skillcheck standard-library-only and test malformed metadata, stale prompts, missing targets, cycles, unreachable skills, and unfinished placeholders. Update its canonical topology expectation in the same change as an intentional handbook route change.

Forward-test a complex or substantially revised skill with a fresh agent when that can be done safely. Give it the skill path and a realistic task, not the intended answer or prior diagnosis, then revise the skill if it cannot guide the task without leaked context.