Author and optimize AI skill definitions with depth validation

Provides a single canonical, 7-step workflow for creating or improving a skill, loading only the specific reference file needed for each sub-task rather than one monolithic guide. Step 1 resolves the target skill path and operation (create, update, synthesize, iterate), classifies the skill (workflow-process, integration-documentation, security-review, skill-authoring, or generic), and asks one direct question only if the class or depth requirement is genuinely ambiguous. Step 2 runs source synthesis when needed - collecting and scoring sources with provenance, applying trust and safety rules to external content, and enforcing depth gates before authoring can start. Step 3 runs an iteration pass first when improving from real outcomes or examples - capturing and anonymizing examples with provenance, re-evaluating behavior against working and holdout slices, and carrying concrete behavior deltas into authoring; it is skipped when the operation isn't iterate. Step 4 authors or updates the actual SKILL.md and supporting files in imperative voice with a trigger-rich description, creating focused reference files or scripts only when justified, and for authoring or generator skills including transformed examples covering the happy path, a secure/robust variant, and an anti-pattern with its corrected version. Step 5 optimizes the skill's description against should-trigger and should-not-trigger query sets to cut false positives and negatives, keeping trigger language generic across Codex and Claude. Step 6 evaluates outcomes with a lightweight qualitative check by default, adding a concise depth rubric for integration/documentation and skill-authoring skills, and running a deeper quantitative baseline-vs-with-skill evaluation only when requested or risk warrants it. Step 7 registers the skill and runs validation with strict depth gates that reject shallow outputs or missing required artifacts.

Use it as the canonical workflow any time a skill is being created, updated, synthesized from sources, or iterated on based on real usage examples - its stated success condition is maximizing high-value input coverage before authoring so the resulting skill has minimal blind spots. Load only the reference path each step actually needs (mode selection, design principles, skill patterns, workflow patterns, output patterns, synthesis path, authoring path, description optimization, iteration path, evaluation path, registration validation) rather than reading everything up front. It is not a substitute for environment-specific validation, testing, or expert review, and its own guidance says to stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

Input is the target skill path, the intended operation, and, when relevant, source material to synthesize or real usage examples to iterate from. Output is a fixed four-part report: Summary, Changes Made, Validation Results, and Open Gaps.

Organized entirely around its own reference-file library (references/mode-selection.md, references/synthesis-path.md, references/authoring-path.md, references/description-optimization.md, references/iteration-path.md, references/evaluation-path.md, references/registration-validation.md, plus pattern references and example profiles under references/examples/); it has no external service integrations beyond whatever repository registration step the target skill library uses.

Skill authors and maintainers who want one consistent, gated process for writing a new skill or improving an existing one from real feedback - covering synthesis, authoring, description tuning, evaluation, and registration - instead of an ad hoc, undocumented process.