| name | generate-skills |
| description | Generate .claude/skills/ definitions โ the capability layer of the workflow system. Identifies recurring workflows and creates SKILL.md files that agents execute and workflows compose. Invoked by scaffold or standalone via /generate-skills. |
Generate Skills
Produce .claude/skills/<name>/SKILL.md files for recurring workflows identified during the grill.
Role in the System
Skills are the capabilities โ the procedures that agents execute and workflows compose into pipelines. Each skill is a reusable building block. Agents use skills to accomplish their responsibilities. Workflows chain skills into end-to-end automation.
Skills (capabilities)
โ Identified from grill decisions + workflow design
โ Executed by Agents (specialist knowledge)
โ Composed by Workflows (pipeline steps)
โ Triggered by Users (slash commands)
Input
Scan results, grill decisions, and the workflow design that determines what capabilities the system needs. Read .claude/scaffold-decisions.md if it exists โ this is the primary source for resolved grill decisions. Skills emerge from:
- Recurring multi-step workflows the user described
- Complex processes that benefit from structured instructions
- Domain-specific tasks requiring specialized knowledge
- Automations the user wants on-demand (not automatic โ those are hooks)
Skill Identification
During the grill, listen for:
- "I always have to..." โ candidate skill
- "The process for X is..." โ candidate skill
- "Claude keeps getting X wrong" โ candidate skill (with corrective instructions)
- "We do X differently than most" โ candidate skill (domain-specific)
Output Format
Each skill at .claude/skills/<name>/SKILL.md:
---
name: <kebab-case-name>
description: <what the skill does and when to use it โ max 1024 chars>
---
Body follows progressive disclosure:
- Metadata (frontmatter) โ always loaded, used for skill discovery
- Body โ loaded when skill is triggered
- Bundled resources โ loaded on demand via file references
Skill Writing Rules
- Imperative/infinitive voice ("Run the test suite" not "You should run the test suite")
- No second person ("Run X" not "You run X")
- Concrete, actionable steps โ not philosophy
- Include examples when the pattern isn't obvious
- Reference bundled files with relative paths for large content
- Add
disable-model-invocation: true for skills with side effects (deploy, publish, etc.)
- Use
$ARGUMENTS for skills that accept input
Common Skill Patterns
| Pattern | Example | Triggers |
|---|
| Verify/validate | /verify โ run all checks before commit | User says "before I commit I always..." |
| Deploy | /deploy-staging โ deploy to staging env | User describes deployment process |
| Review | /review โ structured code review checklist | User has specific review criteria |
| Generate | /generate-component โ scaffold new component | User has boilerplate patterns |
| Debug | /debug-<system> โ systematic debugging for specific system | User describes recurring debug patterns |
Don't Create Skills For
- One-off tasks
- Standard practices well-documented elsewhere
- Project-specific conventions (those go in CLAUDE.md or rules)
- Simple commands (those are aliases, not skills)
Standalone Invocation
/generate-skills โ scan + grill + generate skills.
$ARGUMENTS:
--list โ List candidate skills without generating
--name=verify,deploy โ Generate only specified skills
--quick-grill โ Abbreviated interrogation (3-5 questions)