| name | skill-creator |
| description | Structured workflow for creating or updating skills — from discovery through validation and audit. Produces structurally correct, behaviorally effective skill packages (SKILL.md + companions) using deep-plan-style discovery with skill-specific question categories and convergence-loop verification of codebase claims. |
| compatibility | agentskills-v1 |
| metadata | {"version":3.4} |
Skill Creator
Structured creation workflow that produces high-quality skills by front-loading
decision-making through exhaustive discovery. Effectiveness is measured against
the quality checklist in SKILL_STANDARDS.md.
Critical Rules (MUST follow)
- Discovery before writing — MUST complete discovery and planning before
writing any skill files. No skipping ahead.
- Approval gate before scaffold — MUST present planning output to user and
get explicit approval before creating files.
- Skill-audit is mandatory — MUST run
/skill-audit on every created skill
before considering it complete.
- Under 300 lines — created SKILL.md MUST stay under 300 lines. Extract to
REFERENCE.md if approaching the limit.
- MUST/SHOULD/MAY on every instruction — every instruction in created
skills MUST be classifiable as MUST, SHOULD, or MAY.
- Persist state incrementally — save to state file after every phase. Long
creation sessions WILL hit compaction.
- Recommend with rationale — at every decision point, state your
recommendation and explain WHY.
When to Use
- Creating a new skill from scratch
- Updating or improving an existing skill (major changes)
- User explicitly invokes
/skill-creator
When NOT to Use
- Auditing an existing skill's quality — use
/skill-audit
- Creating an audit-type skill — use
/create-audit
- Checking all skills for issues — use
/skill-ecosystem-audit
- Quick structural check — use
npm run skills:validate
What This Skill Does NOT Do
- Ecosystem-wide analysis (use
/skill-ecosystem-audit)
- Structural validation only (use
npm run skills:validate)
- Behavioral quality evaluation (use
/skill-audit)
- Audit-specific scaffolding (use
/create-audit)
Routing Guide
| Situation | Use | Why |
|---|
| Create new skill | /skill-creator | Full discovery + planning |
| Improve existing skill | /skill-audit first | Audit identifies what to fix |
| Create audit-type skill | /create-audit | Specialized audit scaffolding |
| Check all skills | /skill-ecosystem-audit | Ecosystem-wide analysis |
| Quick structural check | npm run skills:validate | Fast, no AI needed |
Input
Argument: Skill name or description, passed as /skill-creator <name> or in
conversation context.
Output location: .claude/skills/<skill-name>/ (MUST — user-created skills
go here, not in vendor directories).
Complexity Tiers
| Tier | Characteristics | Process |
|---|
| Simple | Reference-only, <50 lines, no companions | Skip discovery, write directly |
| Standard | Workflow, 50-200 lines, 0-2 companions | Full process |
| Complex | Multi-phase, 200-300 lines, 3+ companions | Full process + enhanced planning |
Process Overview
WARM-UP Orientation -> Skill name, process, effort estimate, welcome
PHASE 1 Context -> ROADMAP check, neighbor scan, existing patterns
PHASE 2 Discovery -> Exhaustive questions (6 categories, floor ~12)
PHASE 3 Planning -> Decision record + skill structure plan + approval
PHASE 4 Build -> Scaffold, write SKILL.md + companions
PHASE 5 Validate -> Structural + behavioral checklist
PHASE 6 Audit -> Run /skill-audit on created skill
PHASE 7 Closure -> Retro, invocation tracking, session cleanup
Use phase markers: ======== PHASE N: [NAME] ========
Warm-Up (MUST)
Present before any work: skill name, process overview, complexity tier estimate,
effort estimate (Simple: 10min, Standard: 30min, Complex: 60min+).
"Before we start: any specific concerns, constraints, or existing patterns you
want this skill to follow?"
Done when: User confirms proceed.
Phase 1: Context Gathering (MUST)
- Check ROADMAP.md — verify skill aligns with project direction (MUST). If
misaligned: present conflict, offer proceed/reframe/abort. Do NOT silently
proceed.
- Scan
.claude/skills/ for existing neighbors (MUST)
- Check CLAUDE.md for relevant conventions (MUST)
- Extraction context (MUST): Scan
.research/EXTRACTIONS.md for candidates
relevant to the skill being created (human-readable, grouped by source). For
targeted filtering, query .research/extraction-journal.jsonl (match by
type, tags in notes, or source domain). If matches found, present as "Prior
art from analyzed sources" — these are patterns, architectures, and design
principles already identified from external repos/websites that may inform
the skill's design.
- Read existing skill files if updating (MUST for updates)
- Domain research check (SHOULD) — if the skill being created covers a
domain the AI lacks expertise in (unfamiliar technology, specialized
workflow, external ecosystem), suggest: "This skill targets a domain that
would benefit from research. Run
/deep-research first to understand the
landscape?" Prior research at .research/<topic-slug>/ feeds directly into
better discovery questions and informed defaults in Phase 2.
Reframe path: If context reveals the user doesn't need a skill — they need a
hook, script, CLAUDE.md entry, or agent — present the reframe before proceeding.
Done when: Context gathered, no ROADMAP conflict (or conflict resolved).
Phase 2: Discovery (MUST for Standard/Complex, skip for Simple)
Read .claude/skills/skill-creator/REFERENCE.md for the 6 question categories
with example questions.
Discovery Rules (MUST follow)
- Floor of ~12 questions, no ceiling — ask until ambiguity is resolved
- Front-load critical decisions — scope, architecture, neighbors first
- Offer defaults for every question — "I recommend X because Y. Override?"
- Reference existing patterns — cite actual codebase skills, not generics
- Batch 4-6 related questions — group by category, critical first
- Inter-batch synthesis — summarize learnings before next batch
- State inferences explicitly — "Based on Q3, inferring X for Q7"
- Save decisions with rationale after every batch (MUST) — persist to state
file, include WHY per decision
- Show progress — "Batch 2 of ~3. N decisions captured."
- Multi-agent exploration (SHOULD for Complex) — dispatch Explore agents
to scan the codebase for relevant patterns before presenting questions.
Findings inform defaults and recommendations.
- Cross-batch revision (MUST) — if a later batch answer conflicts with an
earlier decision, present the conflict and allow revision. Update state file
with the revised decision.
- Self-audit design (MUST for Standard/Complex) — discovery MUST include
questions about the created skill's self-audit phase. See REFERENCE.md
Category 2 for the self-audit design questions.
Delegation: If user says "you decide," accept recommended defaults. Document
each delegated decision with rationale.
Contradiction detection: If answers conflict, flag immediately: "Decision N
said X, but this implies Y. Which takes priority?"
Mid-discovery check (MUST after batch 2): "Discovery progress: N questions,
M decisions. Estimated ~K more. Continue or scope-reduce?"
Post-discovery prompt (MUST): "Discovery complete. Anything the questions
didn't cover, or concerns before we plan?"
Done when: All 6 categories covered, no remaining ambiguity, user confirms
nothing missing.
Phase 3: Planning + Approval (MUST)
- Compile decision record — one row per decision, choice + rationale (MUST)
- Present skill structure plan (MUST):
- Files to create (SKILL.md, REFERENCE.md, scripts/, etc.)
- Section outline for SKILL.md
- Guard rails and failure modes
- Integration surface (neighbors, handoffs)
- Compaction resilience design
- Self-audit phase design (MUST for Standard/Complex) — list expected
output artifacts to verify, verification agents to dispatch, orphan
detection scope, and downstream contract checks. Reference
SKILL_STANDARDS.md "Self-Audit at Completion" for the 9 dimensions.
- Present for approval (MUST): "Here's the planned structure. Proceed? [approve
/ approve with changes / reject]"
Revision: If writing later reveals a planning decision was wrong, pause,
note the revision, confirm with user before continuing.
Done when: User approves plan.
Phase 4: Build (MUST)
4.1 Input Validation (MUST)
Validate skill name: lowercase-kebab-case, no spaces, max 40 chars. Check
.claude/skills/ for existing skill. If exists and malformed, offer repair or
fresh start.
4.2 Scaffold
Run init_skill.py <name> --path .claude/skills for new skills. Skip for
updates.
4.3 Write SKILL.md
Read REFERENCE.md for the full content checklist (MUST/SHOULD items).
Structure for attention management:
- Top third (1-100): Critical rules, MUST behaviors, process overview
- Middle third (100-200): Step-by-step procedures, checklists
- Bottom third (200-300): Guard rails, compaction, version history
Point-of-use reminders (MUST):
- Repeat "under 300 lines" in the validation phase
- Repeat "MUST/SHOULD/MAY" where instructions are written
- Repeat "front-load critical rules" in the writing phase
Project conventions: Reference CLAUDE.md by path. Do NOT duplicate
conventions into the skill.
Data standards: If skill produces data files, SHOULD use JSONL as canonical
format with .md for human readability.
Convergence-loop integration (SHOULD): If Phase 2 discovery identified a
discovery or verification phase in the skill being created (T25 question), wire
convergence-loop programmatic mode into that phase. Reference
/convergence-loop SKILL.md "Programmatic Mode" section for the integration
contract.
Self-audit phase (MUST for Standard/Complex): Two parts.
- Prose phase in SKILL.md — write a self-audit phase as the penultimate
phase (before Closure). Reference SKILL_STANDARDS.md "Self-Audit at
Completion" for the 9 verification dimensions. Customize with the skill's
specific deliverables, downstream contracts, and regression detection needs
from Phase 3 planning.
- Scaffold the script — create
scripts/skills/<skill-name>/self-audit.js
per SELF_AUDIT_PATTERN.md. Use
scripts/skills/skill-audit/self-audit.js as the template and replace
domain-specific dimension checks. The Phase 5 prose in SKILL.md MUST invoke
this script first (per pattern doc §Wiring to SKILL.md Phase 5), with prose
handling judgment-only checks afterwards.
If skill is Simple tier (<50 lines, reference-only), skip both parts — the "Done
when:" gate is sufficient per SKILL_STANDARDS.md.
4.4 Write Companions (SHOULD)
Extract to REFERENCE.md if SKILL.md approaches 300 lines. Move examples,
templates, question banks to companion files.
Done when: All files written, content checklist addressed.
Phase 5: Validate (MUST)
-
Run npm run skills:validate (MUST) — structural check
-
Walk the content checklist from REFERENCE.md (MUST) — for each unchecked
item, either address it or document why not applicable
-
Convergence-loop verify (MUST for Complex, SHOULD for Standard) — verify
created skill's codebase claims (neighbors, integration points, existing
patterns) via convergence-loop quick preset. Skip if skill makes no codebase
claims. If claims are wrong: fix in created skill files (Phase 4
re-entry) before continuing validation.
-
Verify line count under 300 (MUST)
-
Verify cross-references resolve (SHOULD)
-
Evidence-based self-check (MUST) — re-read all created files. For each
planning decision from Phase 3, grep the output file for a keyword proving
implementation. Decisions with no grep match are MISSING — fix before
proceeding to Phase 6. See SKILL_STANDARDS.md "Self-Audit at Completion."
-
Self-audit phase present (MUST for Standard/Complex) — verify both: (a)
SKILL.md includes a self-audit phase positioned as penultimate, referencing
SKILL_STANDARDS.md dimensions appropriate to tier; (b)
scripts/skills/<name>/self-audit.js exists and follows
SELF_AUDIT_PATTERN.md (matching CLI
shape, exit codes, SUMMARY block, required helpers).
-
Contract verification wired (MUST if downstream consumers) — verify the
self-audit checks output against documented consumer expectations
Done when: skills:validate passes, behavioral checklist addressed, under
300 lines, codebase claims verified, all decisions grep-verified, self-audit
phase present (Standard/Complex).
Phase 6: Audit (MUST)
Run /skill-audit <skill-name> on the created skill. The audit verifies
behavioral quality that structural validation cannot catch.
Done when: Skill-audit complete, findings addressed.
Phase 7: Closure (MUST)
- Closure signal (MUST): List all created/modified files with paths
- Retro prompt (MUST): "What did discovery surface that you didn't expect?
What did the checklist catch? What would you do differently?" Capture in
state file
process_feedback.
- Session cleanup (SHOULD): Commit created skill files. State file retained
as creation record. Run
/session-end if ending session.
- Version history (MUST): WHAT changed + WHY in every entry.
Guard Rails
- Contradiction: Flag conflicting requirements immediately
- Scope explosion: If planned skill exceeds 300 lines + 3 companions + 5
scripts, pause: "This is growing large. Split into multiple skills?"
- Phase gates: MUST complete Phase 1 before Phase 2. MUST have approved plan
(Phase 3) before Phase 4.
- Disengagement: If user stops mid-skill, save state, list files created,
offer cleanup (delete partial work or keep for later)
Anti-Patterns
Read REFERENCE.md "Anti-Patterns" section for the full 12-item list of common
skill creation mistakes (MUST check during Phase 4 Build).
Compaction Resilience
- State file:
.claude/state/skill-creator.state.json
- Update: After every phase boundary and every discovery batch
- Contents: Skill name, discovery decisions, planning checklist, current
phase, files created, process feedback
- Recovery: Re-invoke
/skill-creator <skill-name> to read state and resume
- Retention: State file retained after completion as creation record and
artifact contract for
/skill-audit handoff
Integration
- Neighbors:
/skill-audit (audit created skill), /skill-ecosystem-audit
(check all skills), /create-audit (audit-specific scaffolding),
/convergence-loop (Phase 4.3: wire into created skill if T25 applies; Phase
5: verify created skill's codebase claims)
- References: SKILL_STANDARDS.md (Phase 5
content checklist), SELF_AUDIT_PATTERN.md
(Phase 4.3 scaffold + Phase 5 verification),
SKILL_AGENT_POLICY.md
- Handoff: State file documents intent + discovery decisions for skill-audit
Version History
| Version | Date | Description |
|---|
| 3.4 | 2026-04-14 | Phase 4.3 + 5.7: scaffold + verify scripts/skills//self-audit.js per pattern |
| 3.3 | 2026-04-04 | Add mandatory self-audit phase for Standard/Complex skills (9 dimensions) |
| 3.2 | 2026-03-15 | Skill-audit: CL verify MUST for Complex, reorder Phase 5, anti-patterns to REFERENCE |
| 3.1 | 2026-03-15 | Convergence-loop integration: Phase 4.3 build + Phase 5 verify |
| 3.0 | 2026-03-08 | Full rewrite from 52-decision audit (48/100 -> target 82/100) |
| 2.2 | 2026-03-07 | SC-1..5: operational deps, compaction MUST, scope, recommendations |
| 2.1 | 2026-03-06 | Add interactive design, verification phase, UX |
| 2.0 | 2026-02-28 | Add attention management, behavioral quality, guards |
| 1.0 | 2026-02-25 | Initial implementation (Anthropic skill, Apache 2.0) |