define-skill

Part of the ACED plugin — see that page for install instructions.

Trigger: “create a skill”, “write a skill for X”, “I want a skill that…”, “turn this workflow into a skill”, “help me build this skill” — even without the word “skill”

Creates or improves a workflow skill — a process, tool-based, or standard SKILL.md — then hands it to the ACED eval loop to spec and score.

What it does

1. Routes away requests that belong to a sibling skill: agents/personas → define-agent, reference-only rule sets → define-governance, extracting the current session → skillify, scoring/adding cases → run / add-scenario, diagnosing failing evals → improve.
2. Loads the skill-design governance for authoring rules.
3. Settles scope, trigger condition, output contract, quality bar, and out-of-scope before writing — asks the user to resolve anything that can’t be inferred.
4. Chooses a pattern (process / tool-based / standard) and a placement (user-global / project-private / project-public), scaffolds the file, and creates a symlink per targeted runtime (Claude Code, Cursor, Codex).
5. Runs the structural audit (cyberplace audit validate) and fixes any CRITICAL or HIGH finding before presenting the skill.

Three entry points

• Impl-producer (dispatched by the SDD conductor against a frozen .feature) — co-produces the SKILL.md and its verification (the frozen .feature’s scenarios plus the node’s eval.md run policy) that run will score.
• Standalone — the user invokes it directly, outside any change request. Scaffolds the skill, then offers the ACED eval loop rather than assuming it.
• Escaped — invoked directly after the request resolves ignored (before any change request opens). Scaffolds, audits, reports, and stops — no eval loop.

Next step

Standalone or impl-producer entries: run sdd:start-mission to author the skill’s .feature and eval suite, or use add-scenario / run to grow and score a golden set.