Skip to content

README.md

Latest commit

 

History

History
105 lines (72 loc) · 5.38 KB

README.md

File metadata and controls

105 lines (72 loc) · 5.38 KB

CraftKit

License: MIT

CraftKit is a cross-agent toolkit for creating, improving, and operationalizing prompts and skills for coding agents such as Claude Code and Codex.

Why CraftKit

Prompt assets and agent skills often become fragmented, provider-specific, and hard to reuse. CraftKit exists to keep them file-first, portable, reviewable, and easy to improve over time.

Install

All eight skills install as Claude Code custom slash commands.

Via npx skills

npx skills add sungjunlee/craftkit

Add -g -y for global install without prompts:

npx skills add sungjunlee/craftkit -g -y

Via Claude Code Plugin Marketplace

/plugin marketplace add https://github.com/sungjunlee/craftkit.git
/plugin install craftkit@craftkit
Install from a local clone 
git clone https://github.com/sungjunlee/craftkit.git
cd craftkit
npx skills add . -g -y

For Codex or any other agent, see Use in other agents below.

The eight skills

Skill Use when
craft-prompt a new prompt is needed from scratch for any LLM (Claude, GPT, Gemini, Perplexity, etc.)
craft-skill-spec a new skill needs a concrete spec based on current CraftKit skill-radar judgments before writing SKILL.md
craft-scaffold a rough idea needs structure — goals, inputs, workflow, outputs — before implementation
craft-critique a prompt or skill "feels off" and a diagnostic pass should come before any rewrite
craft-tune an existing prompt is close but needs targeted, minimal-diff sharpening
craft-survey a new skill should be grounded in prior art, extracting only patterns that carry their weight
craft-autoresearch a prompt or skill works "sometimes" and needs eval-driven iteration
craft-handoff a session is ending and the next session needs a copy-paste-ready continuation prompt (clipboard + optional auto-load on /clear)

Each skill lives at skills/<skill-name>/SKILL.md — plain markdown with YAML frontmatter, loadable as a Claude Code skill or copy-pasteable into any other agent.

Status

Six of the eight skills (craft-prompt, craft-scaffold, craft-critique, craft-tune, craft-survey, craft-autoresearch) have been optimized through craft-autoresearch passes against eval suites — including craft-autoresearch itself (reflexive meta-pass). craft-handoff and craft-skill-spec are new and have not yet been through an autoresearch pass. Per-session baseline → kept-state scores and mutation rationale live in the commit bodies. Run artifacts are preserved at ~/.craftkit/autoresearch/<skill>/<date-slug>/ outside the repo.

What belongs in CraftKit

  • generating new prompts from scratch (task, research, session handoff, templates)
  • prompt design and restructuring
  • reusable skill design
  • diagnostic critique and quality checks
  • iterative improvement loops
  • survey-backed best practices
  • time-aware curation of evolving skill-authoring patterns
  • copy-pasteable outputs for agent workflows

Design principles

  1. File-first and diff-friendly
  2. Small composable units
  3. Explicit inputs and outputs
  4. Cross-agent portability (core skill spines stay provider-neutral; platform-specific detail stays in sub-skills like craft-prompt/guides/)
  5. Eval-driven improvement when possible
  6. Copy-pasteable results over fancy abstractions

For evolving skill-authoring guidance, the craft-skill-spec skill carries its own radar layer at skills/craft-skill-spec/references/radar/ — start with current.md there and consult the dated snapshots only when a watch item needs deeper context.

Use in other agents

CraftKit skills are plain markdown with YAML frontmatter, so they port easily:

  1. Open the relevant SKILL.md.
  2. Paste the body (everything after the frontmatter) into the target agent's system prompt or instructions.
  3. Keep the frontmatter description line as context so the agent knows when to apply the skill.

See docs/examples/tune-a-prompt.md for a walk-through of critiquing an existing prompt, tuning it, and running a short improvement loop.

Prior art

  • sungjunlee/prompt-builder — predecessor project. Its mature prompt-authoring asset (5-step process, 6 building blocks, platform guides, templates) was absorbed wholesale into craft-prompt. Kept on GitHub for reference; new work happens here.
  • karpathy/autoresearch — Andrej Karpathy's ML training-loop project that introduced the autoresearch methodology (give an agent a baseline, let it experiment overnight, keep what improves, discard what doesn't). craft-autoresearch adapts that loop discipline to prompt and skill artifacts instead of model training code.
  • byungjunjang/jangpm-meta-skills — four-skill meta toolkit for Claude Code and Codex (blueprint, deep-dive, reflect, autoresearch). Its autoresearch skill contributed implementation patterns — experiment contract shape, the three-eval-type taxonomy (binary / comparative / fidelity), deletion discipline — that craft-autoresearch builds on.

License

MIT — see LICENSE.