Skip to main content
skill CX Codex

create-skill · sethgammon — Claude Code skill: what it does, how to install

create-skill interviews you about a recurring workflow, compiles it into a structured SKILL.md file, and tests it against real project files.

MIT ★ 937 updated 2026-10-29

Install into Claude Code · Plugin marketplace All methods and platforms ↓
/plugin marketplace add SethGammon/Citadel
/plugin install citadel@citadel-local
Open on GitHub

What create-skill does

The skill turns manual repetitive habits into clear agent instructions. It begins by asking three questions: what steps you repeat, what mistakes usually happen by hand, and what finished output looks like. From these responses, it extracts triggering phrases, boundary conditions, and explicit numbered steps.

The generated instructions land in .claude/skills/{name}/SKILL.md under 500 lines. The tool registers the new entry in .claude/harness.json, updates CLAUDE.md or AGENTS.md, and runs a validation pass against an actual target in the project. If the new workflow yields thin results or misses context, it refines the protocol before handing it over.

This skill is not meant for one-off tasks: it only builds reusable workflows from scratch. If you need to tweak an existing protocol, edit the markdown file directly.

What it helps with

  • turn repetitive manual coding workflows into reusable agent skills
  • define concrete quality gates and failure guards for routine tasks
  • generate and register SKILL.md inside project instruction files
  • test newly authored instructions on actual project files immediately

Example prompts for the agent

  • /create-skill
  • Create a skill for reviewing database migrations before pushing
  • I keep running this manual cleanup pattern, make a skill for it

How to install create-skill

1Plugin marketplace

/plugin marketplace add SethGammon/Citadel
/plugin install citadel@citadel-local

This installs the whole citadel plugin with all its skills

2Skills CLI (Vercel)

npx skills add SethGammon/Citadel --skill create-skill -a claude-code

3Manually: put the files in place

~/.claude/skills/create-skill/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/create-skill/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about create-skill

How to install create-skill in Claude Code?

Add the marketplace using /plugin marketplace add SethGammon/Citadel, then run /plugin install citadel@citadel-local.

What does create-skill do?

It conducts an interview about your routine task, writes a tested .claude/skills/{name}/SKILL.md file, and registers it in your project config.

Can create-skill modify existing skills?

No, it is designed strictly for new workflows. To modify an existing skill, update its markdown file directly.

Source text

SKILL.md · MIT
---
name: create-skill
license: MIT
description: >-
  Creates new skills from the user's repeating patterns. Interview-driven:
  discovers the task, analyzes failure modes, generates a production SKILL.md,
  installs it, tests it on a real target, and teaches the user how to use it.
  Use when a user wants to encode a repeating workflow; do NOT use for one-off
  tasks or modifying existing skills.
user-invocable: true
auto-trigger: false
trigger_keywords:
  - create skill
  - new skill
  - make a skill
  - teach the harness
  - custom skill
  - my own skill
  - skill for
  - automate this pattern
  - repeated pattern
---

# /create-skill — Skill Creator

## Orientation

**Use when:**
- The user says "I keep doing this same thing" or "automate this for me"
- The user wants to encode a workflow they have refined through repetition
- The user says "create a skill for X" or "make a skill that does Y"

**Do NOT use when:**
- The user wants a one-off task done (just do it)
- The pattern has only happened once
- An existing skill already covers this
- The user wants to modify an existing skill (edit directly)

**Output:** A complete `.claude/skills/{name}/SKILL.md` in the project directory, tested and working.

## Protocol

### Step 1: DISCOVER — The Three Questions

Ask these three questions and wait for answers before proceeding.

**Q1: "What do you keep repeating?"**
Listen for: trigger, steps in order, scope, frequency. If vague, probe: "Walk me through the last time you did this."

**Q2: "What mistakes happen when you do it manually?"**
Listen for: forgotten steps, ordering mistakes, convention drift, edge cases. These become guardrails and quality gates.

**Q3: "What does 'done right' look like?"**
Listen for: observable outputs, quality signals, anti-patterns. These become quality gates and exit protocol.

### Step 2: ANALYZE — Extract the Skill's DNA

Internal working material — do not show to user.

**2a. Identity statement:** "You are a {role} that {does what} to ensure {outcome}." Must distinguish this skill from all others.

**2b. Trigger keywords (5-10):** Specific enough to avoid false matches. Check existing `.claude/skills/` for conflicts.

**2c. Protocol steps:** Transform "what I do" into numbered steps where:
- Each step has a clear input and output
- Decision points have explicit criteria ("IF x THEN y, ELSE z")
- Steps reference concrete things (file paths, commands, patterns)

Bad: "3. Review the code for issues."
Good: "3. Read every function. For each, check: (a) return type explicit, (b) error cases handled, (c) no input mutations. List violations with line numbers."

**2d. Quality gates:** Yes/no verifiable questions, no subjective criteria. Map each to a user "done right" criterion or common mistake.

**2e. Pitfalls:** From Q2. Become warnings or guard clauses in the protocol.

### Step 3: GENERATE — Write the SKILL.md

Every section required. File MUST be under 500 lines.

```markdown
---
name: {kebab-case-name}
description: >-
  {One to three sentences. Start with verb. Use when X; do NOT use for Y.}
user-invocable: true
auto-trigger: false
trigger_keywords:
  - {keyword 1}
  - {keyword 2}
---

# /{name} — {Readable Title}

## Orientation

**Use when:**
- {condition}

**Do NOT use when:**
- {exclusion}

**What this skill needs:**
- {required input}

## Protocol

### Step 1: {VERB — Name}
{Exact instructions. What to read, what to produce. No vague directives.}

### Step 2: {VERB — Name}
{Continue...}

## Quality Gates

- [ ] {Verifiable gate}

## Exit Protocol

{Exact output format}
```

**Writing rules:**
1. Steps must be reproducible by a different AI session with no memory of this conversation.
2. No hedge language — "do X" not "consider X".
3. No filler — delete sections that add no new information.
4. Include examples only where the pattern is non-obvious.
5. Encode the user's taste: "I hate when it does Z" → explicit prohibition.

### Step 4: INSTALL & REGISTER

1. Create `.claude/skills/{name}/`
2. Write the SKILL.md
3. Verify file exists and is readable
4. Register with router:
   a. Read `.claude/harness.json`, Citadel's shared config path for all runtimes including Codex (create with `{}` if missing)
   b. Add skill name to `registeredSkills` array
   c. Update `registeredSkillCount` to match array length
5. Only add the skill to an active project instruction file that already has an explicit skills section. Codex uses `AGENTS.md`. Claude Code uses `CLAUDE.md` when present; with confirmed `AGENTS.md` support, it uses `AGENTS.md` by default if no project `CLAUDE.md` exists, or both when configured under **Project instructions**. Check the active setting before editing and preserve user-authored guidance.
6. Run `/reload-skills` if the running Claude Code version supports it; the skill is live without a session restart. If unsupported, tell the user a fresh session is required before `/{name}` resolves.

### Step 5: VERIFY — Test on a Real Target

1. Find a real target in the current project the skill applies to.
2. Run the skill's protocol following the SKILL.md exactly — pretend you are a different AI session with no memory of this conversation.
3. Evaluate: did every step have enough information? Did any step require missing context? Did quality gates catch real issues?
4. If it fails or produces thin output, fix SKILL.md (common fixes: step too vague, step assumed context, gate not checkable, exit output missing fields).
5. Run again after fixes. Must pass on second attempt. If it fails twice, discuss scope with user.

### Step 6: TEACH — Explain What Was Built

**A. How to invoke:**
- Direct: `/{name}` or `/{name} [target]`
- Via router: `/do {natural language matching trigger keywords}`

**B. How it works (30-second version):**
- Steps summary, quality gates summary, exit output summary

**C. How to modify:**
- File: `.claude/skills/{name}/SKILL.md`
- Add step: new `### Step N` in Protocol
- Change quality standards: edit Quality Gates checkboxes
- Change trigger words: edit `trigger_keywords` in frontmatter
- Split if too large: create two skills, move steps between them

## Contextual Gates

**Disclosure:** "Creating skill '{name}'. Will create `skills/{name}/SKILL.md`."
**Reversibility:** green — creates `skills/{name}/SKILL.md` only; undo by deleting the directory.
**Trust gates:**
- Any: full skill creation, install, and teach workflow.

## Quality Gates

- [ ] Three discovery questions asked and answered
- [ ] Generated SKILL.md follows exact format (frontmatter + required sections)
- [ ] Frontmatter has: name, description, user-invocable, trigger_keywords
- [ ] Description includes use-when / do-not-use orientation
- [ ] Protocol steps specific enough for a different AI session to follow
- [ ] No steps contain vague directives ("review", "consider", "ensure quality")
- [ ] Quality gates all yes/no verifiable
- [ ] Trigger keywords do not conflict with existing skills
- [ ] SKILL.md is under 500 lines
- [ ] Skill tested on a real target in the current project
- [ ] Test produced meaningful output
- [ ] User taught invocation, mechanics, and modification
- [ ] File installed at `.claude/skills/{name}/SKILL.md`

## Exit Protocol

```
SKILL CREATED

Name: {name}
Path: .claude/skills/{name}/SKILL.md
Invoke: /{name} [target]
Route via: /do {example natural language}

What it does:
  {One sentence description}

Steps: {N} steps
Quality gates: {N} gates
Lines: {line count}/500

Tested on: {target description}
Test result: PASS

Trigger keywords: {comma-separated list}
Reversibility: green — delete skills/{name}/ to undo
```

```
Try it now: /{name} {suggested first target}
```

© SethGammon · MIT · SethGammon/Citadel

Similar skills

All skills for Agent setup: memory, context, orchestration →

create-skill by other authors comes first.

create-skill NeoLabHQ create-skill — a skill for Claude Code, Codex, Cursor and other agents that walks through writing a new skill using a TDD approach. CX Codex +1 1,300installs create-skill siviter-xyz create-skill is a Claude Code skill for building or revising skills that add capabilities to an agent. It guides requests about skill structure, instructions, and bundled resources. CX Codex +1 2,800installs create-skill aeonfun create-skill turns a one-line description into a finished skill: parses the request, checks for duplicates, researches the APIs and opens a pull request. CX Codex +1 ★ 797repo stars create-skill Samarth Gupta create-skill guides Claude Code through building a skill or slash command by clarifying its purpose and activation scenarios. Use it to draft a new instruction file. CX Codex +1 ★ 276repo stars create-skill thedesignproject This Claude Code skill helps create Cursor skills and explains the structure of SKILL.md. It applies to requests for a new skill or questions about its instruction file. CX Codex +1 ★ 107repo stars create-pr OrchestKit create-pr — a skill that opens a GitHub pull request after running tests, linter, types, and security checks. Triggers on explicit invocation; plain gh pr create stays a command. CX Codex +1 209installs task-observer Eoghan Henn and contributors task-observer — a skill for Claude Code and five other platforms records user corrections and friction during active sessions. Translates recorded observations into skill updates and new playbooks. CX Codex +1 19,300installs qwenchance thananon qwenchance is a skill for long, multi-step work: it detects repeated actions, stops endless second-guessing, and monitors context. It activates when a model gets stuck or nears its context limit. CX Codex +1 2,600installs task-dependency-patterns OrchestKit task-dependency-patterns is a Claude Code skill for splitting complex work into linked tasks. It covers status tracking and coordination across parallel work. CX Codex +1 214installs Tallyfy Workflow Automation tallyfy Tallyfy Workflow Automation is an MCP server for Claude Code that connects an AI assistant to Tallyfy to manage tasks, processes, and templates. It supports workflow operations through a chat-based client. CX Codex +1 ★ 1repo stars jira-automation davila7 jira-automation equips Claude Code to query and manage Jira tasks and sprints via Rube MCP. Useful when filtering issues or organizing backlogs directly from the console. CX Codex +1 50installs

From the same set55

The whole set →
map map generates a JSON index of a repository's files, exports, imports and dependency graph so an agent can locate relevant code without extra Glob or Grep calls. CX Codex +1 ★ 937repo stars phase-validator The phase-validator subagent for Claude Code compares phase handoff reports with exit conditions and returns a pass or fail verdict. It checks automated development pipelines. CX Codex ★ 937repo stars fleet fleet splits a big job into campaigns and runs them in waves of 2–3 agents, handing each wave's findings to the next. Fits when the tasks don't touch the same files. CX Codex ★ 937repo stars workspace workspace coordinates changes across multiple repositories in a single task. Breaks work into waves, runs campaigns per repo, verifies contracts between them. CX Codex +1 ★ 937repo stars arbiter arbiter is a Claude Code subagent that reruns project checks and gives a final verdict on changes. It steps in after a worker runs out of retries or when a run needs a holistic completion review. CX Codex ★ 937repo stars infra-audit infra-audit maps a project's databases, caches and queues from docker-compose, env files and ORM configs, then flags infrastructure layers that look missing. CX Codex +1 ★ 937repo stars refactor Multi-file refactoring with automated Git rollback, typechecking, and test baseline verification across edits. CX Codex +1 ★ 937repo stars review This skill reviews code in five passes, covering correctness, security, performance, readability, and consistency. Use it on a file, directory, or Git diff. CX Codex +1 ★ 937repo stars

Project sponsors

Companies that keep this analytics open