From launchpad
Creates new Claude Code skills using the 7-phase Meta-Skill Forge methodology. Produces quality-validated skills with progressive disclosure, evaluation loops, and contrarian analysis. Use when creating a new skill, teaching Claude a new workflow, building a custom agent capability, or automating a repeatable task. Triggers on: create skill, new skill, build skill, teach Claude, make a skill for.
How this skill is triggered — by the user, by Claude, or both
Slash command
/launchpad:lp-creating-skillsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
NEVER produce a skill that looks like what Claude would generate with no skill loaded. Every skill must demonstrate structural differentiation from baseline output.
NEVER produce a skill that looks like what Claude would generate with no skill loaded. Every skill must demonstrate structural differentiation from baseline output.
Activates when:
Example invocations:
/lp-create-skill frontend development/lp-create-skill "API testing" based on docs/articles/api-guide.md"build a skill for database migrations""teach Claude how to do code reviews"| Request | Use Instead |
|---|---|
| Quick one-off prompts | Direct prompting -- no skill needed |
| Modifying an existing skill | /lp-update-skill |
| Creating cognitive profiles | .claude/profiles/PROFILE-TEMPLATE.md directly |
| Evaluating a skill in isolation | skill-evaluator sub-agent directly |
| Step | Phase | Visible Output |
|---|---|---|
| 1 | Parse input | Skill topic + optional context files identified |
| 2 | Research | Research brief from two-wave sub-agents |
| 3 | Extract | Validated answers from 4 collaborative rounds |
| 4 | Analyze | Contrarian frame with engineering-away plan |
| 5 | Decide | Architecture complexity tier (Simple / Moderate / Full) |
| 6 | Write | SKILL.md + reference files + eval scenarios |
| 7 | Validate | Evaluation report (PASS / FAIL with diagnostics) |
| 8 | Ship | Registered skill + usage guide presented to user |
Every step produces a visible artifact. No step is skipped.
Read references/METHODOLOGY.md for the full two-wave sub-agent protocol before proceeding.
Spawn three sub-agents simultaneously:
After Wave 1 completes, spawn three more:
If the user provided source files ("based on ..."):
Output: Research brief summarizing findings across all sub-agents.
Read references/METHODOLOGY.md for Mode A vs Mode B extraction protocols before proceeding.
Auto-detect based on Phase 1 inputs:
Each round is a focused conversation with the user:
| Round | Focus | Question Type |
|---|---|---|
| 1 | Scope | What does this skill do and NOT do? What triggers it? |
| 2 | Differentiation | What makes this approach different from a generic prompt? |
| 3 | Structure | What phases, reference files, and outputs does the skill produce? |
| 4 | Breaking Points | Where does the skill fail? What edge cases exist? |
Present 3-5 MCQ questions per round. User responds with shorthand (e.g., "1A, 2C, 3B").
If the user defers on a round (e.g., "skip this" or "just decide"), answer the remaining questions using research findings from Phase 1. Skip remaining rounds only if the user explicitly requests it.
Output: Validated extraction answers for all 4 rounds.
Read references/CONTRARIAN-FRAME.md before proceeding.
Write the generic version. Ask: "If Claude received just the topic name and no skill, what would it produce?" Write 3-5 bullet points covering predictable structure, vocabulary, and assumptions.
Name every predictable pattern. List the default sections, ordering, hedge phrases, and assumptions that a skillless Claude would produce.
Challenge 2-3 assumptions. Identify the strongest assumptions in the generic version. For each one, articulate why it fails for this specific use case.
Write the engineering-away plan. For each predictable pattern, state the specific structural or behavioral difference the skill will enforce.
Present the contrarian frame to the user. Two possible outcomes:
Output: Confirmed contrarian frame with engineering-away plan.
Choose complexity based on discoveries from Phases 1-3:
Present the recommendation to the user with this format:
Architecture: [Simple / Moderate / Full]
Rationale: [1-2 sentences explaining why]
File plan:
.claude/skills/<name>/SKILL.md
.claude/skills/<name>/references/FILE-1.md (if Moderate/Full)
.claude/skills/<name>/references/FILE-2.md (if Full)
.claude/skills/<name>/evals/ (if Full)
Get user confirmation before proceeding.
Output: Confirmed architecture tier + file plan.
Read references/SKILL-TEMPLATE.md and references/REFERENCE-TEMPLATE.md before proceeding.
Save files to .claude/skills/<skill-name>/:
SKILL.md -- the orchestratorreferences/*.md -- one file per concern (if Moderate/Full)If the skill was built from user-provided source material (Mode A), spawn a fidelity-check sub-agent (Sonnet, read-only):
Output: Saved skill files + fidelity report (if applicable).
Read references/QUALITY-GATES.md before proceeding.
Spawn a skill-evaluator sub-agent (Sonnet, read-only access to skill files).
The loop:
GENERATE skill files
|
EVALUATE (3 passes: first-principles, baseline detection, Anthropic checklist)
|
ALL PASS? --yes--> Proceed to Phase 7
| no
DIAGNOSE specific failures (explain WHY, not just WHAT)
|
IMPROVE (rewrite targeting diagnosed weaknesses)
|
RE-EVALUATE (max 3 cycles)
If the skill still has failing gates after 3 improvement cycles:
Output: Evaluation report (PASS or FAIL with remaining issues).
Read references/EVAL-TEMPLATE.md before proceeding.
Create at least 3 eval scenarios:
Each scenario includes a baseline comparison (what Claude does WITHOUT the skill).
Save to .claude/skills/<skill-name>/evals/.
Update these files:
docs/skills-catalog/skills-usage.json -- Add "<skill-name>": "YYYY-MM-DD" to the skills object (use today's date). Create the file with initial structure if it doesn't exist: {"last_audit_date": "YYYY-MM-DD", "skills": {}}docs/skills-catalog/skills-index.md -- Add the skill to the correct group in both the Quick Reference table and Detailed Descriptions section. Canonical groups (in order): Design & UI, Frontend Engineering, Backend Engineering, Data & Database, Testing & QA, DevOps & Infrastructure, Security & Auth, API & Integrations, Billing & Payments, Build Pipeline, Quality & Workflow, Documentation, Meta (Skill Management), Other. Assign the next sequential number. Place the skill in the best-fit group. Use "Other" only if no canonical group fits. Add a row to the group's Quick Reference table and a full description entry under the group's Detailed Descriptions heading.Deliver a structured summary:
## Skill Created: <name>
### File Tree
.claude/skills/<name>/
SKILL.md (XXX lines)
references/
FILE-1.md
FILE-2.md
evals/
eval-scenarios.md
### Architecture
[Simple / Moderate / Full] -- [1 sentence rationale]
### Evaluation Findings
[PASS / FAIL] -- [summary of results]
### CLAUDE.md Changes
- Added to workflow commands table: [entry]
- Added to Available Sub-Agents: [entry]
### Usage
Invoke with: /lp-create-skill <topic>
Or say: "create a skill for <topic>"
Ask: "Commit these files, or adjust something first?"
Before delivering the skill to the user, confirm every item:
.claude/skills/<skill-name>/docs/skills-catalog/skills-usage.json updated with new skilldocs/skills-catalog/skills-index.md updated with new skill entryIf any item fails, return to the relevant phase and fix before delivering.
NEVER produce a skill that looks like what Claude would generate with no skill loaded. Every skill must demonstrate structural differentiation from baseline output.
Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub builtform/launchpad --plugin launchpad