craft-skills
A Claude Code plugin providing structured development skills for any project.
Replaces generic AI workflows with battle-tested, structured processes: collaborative brainstorming, planning, parallel agent development, testing, systematic debugging, and self-improvement. Works with any tech stack — frontend, backend, or full-stack.
Installation
/plugin marketplace add alexiolan/craft-skills
/plugin install craft-skills@craft-skills
Updating
/plugin update craft-skills@craft-skills
Note: Claude Code has a known bug where plugin update may report "already at latest" even when a new version exists. If that happens, run this in your terminal first, then retry the update:
git -C ~/.claude/plugins/marketplaces/craft-skills pull --ff-only
Quick Start
Once installed, craft-skills activates automatically. Just describe what you want to build:
/craft Add a user notifications system
What happens:
- Brainstorm — Claude explores your codebase, asks clarifying questions one at a time, proposes 2-3 approaches, and collaborates with you on a design
- Plan — Creates a detailed implementation plan with exact file paths, task ordering, and parallel/sequential dependencies
- Develop — Dispatches parallel developer agents, each implementing one task. Agents share state and coordinate via a shared-state file
- Test — Runs browser tests against the implemented feature
- Report — Summarizes what was built, decisions made, and any open items
You approve or redirect at every phase — nothing runs without your sign-off.
Other entry points:
/implement Add a logout button # Skip brainstorm, go straight to plan → develop
/architect 03 # Plan only (reads .claude/prompts/03-*.md)
/debug The API returns 500 on save # Investigate before fixing — no guessing
/simplify # Review recent changes for quality
/reflect project # Audit your Claude configs
Skills
Pipeline Skills (Craft Variants)
All craft variants share the same pipeline: Brainstorm → Plan → Develop → Test → Report. They differ in which models handle each phase:
| Skill | Who implements | Who reviews | Requirements |
|---|
| craft | Claude (Sonnet) | Claude (Opus/Sonnet) | None |
| craft-ace | Qwen3.6 (local LLM) | Qwen3.6 (review loops) | LM Studio |
| craft-duo | Claude + Codex (GPT-5) | Claude (Opus/Sonnet) | Codex CLI |
| craft-local | Claude (Sonnet) | Claude + Qwen3.6 | LM Studio |
| craft-squad | Claude + Codex (GPT-5) | Claude + Qwen3.6 | Codex CLI + LM Studio |
See Choosing a Variant for detailed comparison with real benchmark data.
Shortcut Skills
| Skill | Description |
|---|
| implement | Fast pipeline: architect → develop → test. For clear, well-understood requirements. |
| finalize | Post-plan pipeline: develop → test. Use when a plan already exists. |
| architect | Create an implementation plan. No code written. |
| develop | Execute an approved plan using parallel agents. |
| browser-test | Run parallel browser-based UI tests. |
Design-Layer Skills
Deterministic UI-quality gates. Invoked automatically by architect/craft/implement when the feature involves UI (no ambient auto-triggering — explicit, gated).
| Skill | Description |
|---|
| aesthetic-direction | One-time-per-project design-language contract written to .claude/aesthetic-direction.md. Auto-generated on first UI feature, refineable by hand. Invokes frontend-design if installed; falls back to observed-tokens template if not. |
| ux-brief | Per-feature UX brief with diagnosis + prioritized patches + success criteria, written to .claude/plans/{feature}/ux-brief.md. Solo mode uses frontend-design alone; combined mode adds ui-ux-pro-max for complex UI (triggered by explicit complexity: frontmatter or inferred from keywords). |
| design-review | Post-develop visual regression + contract audit. Captures screenshots, dispatches Haiku-vision agent against the aesthetic direction and feature brief. Escalates to Sonnet only on MAJOR issues (Haiku-first cost model, ~15-20K tokens per audit). |
Optional external skills (auto-detected, silent skip if not installed):
Process Skills
