skill-authoringlisted

Create focused skill from human-verified workflow. Test behavior against no-skill baseline. ## Scope Use for reusable procedures with multiple steps and judgment. Put one stable rule in `AGENTS.md`. Keep volatile project state in source docs/config. Do not create skill for one-off task or declarative facts. ## Inputs - Human-owned workflow, examples, constraints, and success criteria. - Trigger phrases and negative triggers. - Required inputs, outputs, tools, and target surfaces. - Applicability boundary and lighter fallback. - Existing skill inventory and repo placement rules. Agent may scaffold, organize, compress, and test. Human owns domain truth. [SkillsBench](https://arxiv.org/abs/2602.12670) found self-generated packs below no-skill baselines in tested setups; discovery and creator/solver interference also contributed. Treat result as warning against unreviewed generated procedure, not ban on agent help. ## Workflow ### Step 1: Confirm Need Create skill when workflow repeats, requires judgment, contains reusable traps, or should load only for relevant task. Reject when one rule, source link, or ordinary code/doc solves need. ### Step 2: Bound Skill Write before body: - Name: lowercase kebab-case job. - Trigger: user language and file/domain cues. - Negative trigger: nearest task that should not load it. - Inputs: facts and artifacts required. - Outputs: files or response produced. - Scope: included and excluded work. - Applicability: evidence that justifies