Add a Project-Specific Skill

Add a Project-Specific Skill

You are the Conductor (or a delegated authorized agent) running the end-to-end skill-creation flow. The actual SKILL.md drafting is delegated to the bundled skill-creator (Apache 2.0, at skills/skill-creator/). This wrapper enforces governance around the drafting.

Authorization

Only these agents may invoke this skill:

• chief-ai-officer
• software-architect
• project implementation agents (CAIO-authored from docs/templates/software-engineer-template.md) — task-helper scope only
• tech-lead (escalation-level helpers only)

Reviewers (Senior Software Engineer, QA Engineer), Product Manager, Project Manager, and Technical Writer must escalate to one of the authorized agents.

Invocation

/atelier:add-skill <kebab-skill-name> "<reason>"

• <kebab-skill-name> — the new skill's directory name (e.g., slack-thread-summarizer).
• <reason> — one sentence explaining the capability gap.

Workflow

1. Reuse audit (mandatory)

Run the four-step audit per docs/process/capability-management.md:

1. Inventory — list installed plugins / skills / MCP servers.
2. Search — query Anthropic skills marketplace, public marketplaces, MCP registry, language libraries.
3. Evaluate — does any candidate cover ≥80% of the need?
4. Decide — reuse / extend / create.

Record the audit evidence (candidates considered, why rejected) — this evidence is required for the capability-log entry.

If reuse is possible: stop, use the existing skill, do not create.

2. Invoke the bundled skill-creator

Delegate to /atelier:skill-creator (the Apache-2.0-licensed bundled skill at skills/skill-creator/). Provide:

• The skill name and reason.
• The intended location (almost always <user-project>/.claude/skills/<name>/).
• Any inputs/outputs specification.

skill-creator returns a draft SKILL.md following Anthropic's authoritative structure.

3. Place the file

Project-specific skills go to <user-project>/.claude/skills/<kebab-skill-name>/SKILL.md.

Plugin-level skill contributions (would change atelier itself) are out of scope here — they require a separate plugin PR.

4. Capability-log entry

Append to docs/agents/capability-log.md:

| <YYYY-MM-DD> | skill | <kebab-skill-name> | <invoking-agent> | <reuse audit summary> | <one-sentence justification> | — |

5. User approval (per involvement level)

Involvement	Behavior
Fully Autonomous	Notify via atelier-notify (if configured); proceed.
Milestone Checkpoints (default)	Wait for explicit user approval.
PR-Level / Detailed Supervision	Wait for explicit user approval; show full draft.

6. Reload reminder

Print:

Skill created at .claude/skills/<name>/SKILL.md.
Run `/reload-plugins` to activate it before the next task.

The plugin's UserPromptSubmit hook will also remind the user if the reload is not run within the next session.

Anti-patterns

• Creating a skill whose purpose duplicates an existing one (audit must show "not a fit").
• Inventing a skill name that does not match the kebab-case file name.
• Skipping the capability-log entry.
• Bypassing user approval when involvement level requires it.
• Creating plugin-level skills via this flow (those need a plugin PR, not a project addition).

Reference Documents

• docs/process/capability-management.md
• skills/skill-creator/SKILL.md (bundled drafting engine)