| name | creating-skills |
| description | Use when creating or editing a composable skill within a PAS agent or process. Usually invoked by creating-agents, not directly by users. |
Creating Skills
Create a composable skill following the Agent Skills open standard. Skills define HOW to do a specific thing. They are agent-facing instruction sets, never user-facing. Skills live inside their owning agent or process by default.
Workflow
1. Determine Purpose
Define what the skill does:
- Purpose: what specific capability does this skill provide?
- Consumers: which agent(s) will use this skill?
- Degrees of freedom: where should the agent exercise judgment vs follow strict rules?
- Input: what does the agent need before using this skill?
- Output: what does the skill produce?
2. Check for Overlap
Before creating a new skill:
- Check existing skills within the owning agent (
<marketplace>/plugins/<plugin>/processes/{process}/agents/{agent}/skills/)
- Check existing skills within the process (
<marketplace>/plugins/<plugin>/processes/{process}/)
- Check
${CLAUDE_PLUGIN_ROOT}/library/ for global skills that already do this
- If overlap exists: extend the existing skill or reference it instead of duplicating
3. Apply Granularity Heuristics
Default to one skill (simpler). Split when any heuristic triggers:
| Heuristic | Question | If Yes |
|---|
| Feedback | Can you improve one part without touching the other? | Split into separate skills |
| Reuse | Could another agent use one part but not the other? | Split into separate skills |
| Size | Does the skill exceed ~5000 tokens? | Flag for evaluation: split, restructure, or explicitly justify |
When in doubt, keep it as one skill. You can always split later when feedback indicates the need.
4. Generate the Skill
Run the generation script with all decisions from steps 1-3:
bash ${CLAUDE_SKILL_DIR}/scripts/pas-create-skill \
--process {process-name} \
--agent {agent-name} \
--name {skill-name} \
--description "Use when {triggering conditions}. {What capability this provides.}" \
--overview "{Core principle in 1-2 sentences}" \
--when-to-use "{Specific trigger conditions}" \
--when-not-to-use "{When NOT to use}" \
--step "{Step 1 instruction}" \
--step "{Step 2 instruction}" \
--output-format "{What the skill produces}" \
--quality-check "{Self-check criterion}" \
--common-mistake "{Known pitfall}" \
--base-dir {directory}
Repeatable flags: --step (required, at least one), --quality-check, --common-mistake.
Optional: --base-dir sets the root directory for output (default: current directory).
Key rules from the Agent Skills spec:
- Description = when to use, NOT what it does. Start with "Use when..."
- Progressive disclosure: SKILL.md is the overview. Add heavy material to
references/.
- Concise: only add what Claude doesn't already know. Challenge each paragraph.
- Consistent terminology: pick one term, use it everywhere.
- SKILL.md must be under 500 lines.
5. Library Graduation Check
After creating the skill, check if it should be in library/ instead:
- Is this exact skill already used by another agent in a different process?
- Would a second process/agent benefit from this skill without modification?
If yes to either: move to <marketplace>/plugins/<plugin>/library/{skill-name}/ and reference from both locations. If no: keep it local. Skills start local and graduate to the library only when reuse is proven (used in 2+ places).
