Skill Authoring

Skill Authoring

Inherit the default authoring workflow

Use the installed $skill-creator skill to apply the default authoring workflow. Read it completely and apply its initialization, naming, frontmatter, progressive-disclosure, metadata, validation, and testing guidance.

Apply this skill after $skill-creator. Where the two differ, retain the default requirements and add the repository hierarchy and engineering-contract requirements below.

Build a routed skill hierarchy

Store every discoverable skill as a direct child of .codex/skills/<skill-name>/; do not nest discoverable skill directories inside one another.

Express hierarchy as actionable routing at any depth:

AGENTS.md --Use A to do X--> skill A --Use B to do Y--> skill B --> ...

Write each repo-local route in this form:

Use [skill-name](relative/path/to/SKILL.md) to perform a specific task.

Make the action specific enough that an agent can decide whether to load the linked skill. Put routes to broad skills in AGENTS.md. Put routes to narrower skills in the broad skill that delegates the work. A leaf skill needs no hierarchy boilerplate.

Do not add structural ancestry sections or require a downstream skill to link back to its router. The forward Use ... to ... statements define the hierarchy.

Read the skill routing contract when creating, moving, routing, or auditing a skill.

Treat cross-skill references that do not tell the agent to use a skill for a task as supporting links, not hierarchy routes.

Author project engineering contracts

Capture durable behavior and workflows needed to develop, review, operate, and verify AgentX. Treat the Go source, tests, and skills as cooperating project authorities; update the owning skill when implementation work changes its contract.

Prefer observable, language-neutral contracts:

• responsibility and boundaries
• inputs, outputs, and data shapes
• state and lifecycle transitions
• ordering, concurrency, and timing behavior
• invariants and decision rules
• error, cancellation, retry, and recovery behavior
• user-visible behavior and integration boundaries
• edge cases and acceptance scenarios

Keep contracts focused on behavior and ownership instead of duplicating source listings or private-symbol inventories. Link relevant source or tests when that helps contributors navigate the current repository, but state durable rules in the skill itself.

Put shared concepts in the broadest skill that needs them. Put specialized behavior in the narrowest routed skill that owns it. Do not duplicate the same rule across routing and routed skills.

Derive revisions from a working session

When asked to analyze a session and update skills, first inventory:

• explicit asks and requested changes;
• observed failures, logs, and reproduction commands;
• user corrections and stated operating preferences;
• implementation changes already made during the session.

Classify each item as a durable product contract, reusable troubleshooting workflow, implementation evidence, current conformance gap, or transient artifact. Do not preserve session IDs, timestamps, credentials, temporary paths, one-off outputs, or implementation-library choices unless they are part of a public compatibility contract.

Map each durable item to the narrowest existing owner. Compare its source, tests, current contract, and conformance profile before editing. State desired behavior and acceptance evidence in the owner; when source does not yet satisfy the new requirement, mark that boundary partial or unavailable instead of silently describing it as implemented. Preserve user intent as observable behavior—for example level selection, diagnostic evidence, recovery, and remediation—not as a transcript of the conversation.

Author architecture diagrams

For every diagram-bearing implementation skill, apply the shared architecture diagram contract. Describe the concrete question answered by each asset, identify the authoritative prose, and ensure every page self-identifies its product route, lifecycle position, owned boundary, neighboring owners, edge grammar, contract anchors, and normative limits without relying on its filename.

Update a diagram whenever ownership, topology, lifecycle, ordering, concurrency, durability, trust, or recovery changes. Route fan-out through distinct ports, route feedback around the forward flow, label every behavioral edge, and never use stage numbers for a partial order. Render every changed page at 100% and inspect it; valid XML is not evidence of readable geometry.

For implementation-architecture skills, regenerate shared overviews with generate_drawio.rb, normalize hand-authored pages with enhance_custom_drawio.rb, and run each tool's --check mode. Render every changed page at exact scale with ruby .codex/skills/implementation-conformance-audit/scripts/render_drawio_preview.rb INPUT.drawio --output-dir OUTPUT_DIRECTORY --page PAGE, visually inspect it, and finish with the implementation audit. The enhancer supplies standardized orientation and connector safeguards; the author still owns semantic labels, meaningful lanes, decision branches, and fidelity to prose contracts.

Workflow

1. Read AGENTS.md, inventory .codex/skills/*/SKILL.md, and follow existing Use ... to ... routes relevant to the task.
2. Choose the narrowest existing file that should route to the new or moved skill. Route directly from AGENTS.md only when no existing skill owns the concern.
3. Define the new skill's responsibility so it is cohesive, non-overlapping, and small enough to load independently.
4. Initialize the skill with the default $skill-creator workflow under .codex/skills/<skill-name>/.
5. Write the project engineering contract. Use references only for detailed material that would otherwise bloat SKILL.md.
6. Add a precise Use [skill](path) to ... statement to the selected routing file. Add no backlink solely to represent routing.
7. Add further routes inside the new skill only when it delegates narrower work.
8. Generate or refresh agents/openai.yaml and run the default skill validator.
9. Starting at AGENTS.md, audit all repo-local Use ... to ... routes and repair broken targets, cycles, ambiguous actions, and unreachable skills.
10. Report the new or changed routing chain using repository-relative paths.

Route skills by task

When a task matches a routing statement, read and apply the linked skill. Continue following narrower routing statements only while they match the requested work.

Interpret routed skills as focused workflows:

• Routing skills define shared vocabulary, invariants, and dispatch decisions.
• Routed skills add narrower behavior without repeating unrelated routing context.
• A specialized skill must not silently weaken requirements already applied earlier in the route.

Completion criteria

Finish only when:

• the default skill validator passes for every changed skill
• every repo-local skill is reachable through an actionable route from AGENTS.md
• every Use ... to ... target exists and its action is unambiguous
• no routing chain is cyclic
• no structural ancestry metadata or routing-only backlinks remain
• the authored knowledge gives contributors actionable ownership, behavior, and verification guidance
• examples and acceptance scenarios cover normal behavior and material edge cases
• every diagram is understandable without its filename or surrounding Markdown
• every behavioral edge is labeled and routed without ambiguous overlap
• every page identifies its broader product position, owned/deferred scope, prose anchors, and authority