CraftKit

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.

CraftKit covers two wedges. The first is artifact quality: author prompts, design skills, critique assets, run eval-backed experiments, configure repo harnesses, and carry sessions forward (craft-*). The second is the repo spec axis: spec-charter -> spec-system-map -> spec-grill turns a brownfield repo's direction, system shape, and capability boundaries into reference contracts that other tools can consume — most directly dev-backlog, which measures sprints and triage against them.

CraftKit is not a general coding-agent workflow suite, project-management layer, deployment system, or runtime framework. The spec axis defines what good looks like — it does not manage tasks, sprints, or backlog priority; that stays with dev-backlog. When a workflow needs those things, CraftKit should produce clear files, specs, or handoffs that another tool can use rather than becoming the tool itself.

30-second path

Start with the smallest skill that does the job:

If you need to...	Use
write a new prompt or reusable prompt template	craft-prompt
review an existing prompt or skill without changing it	craft-critique
improve an existing prompt or skill in place	craft-critique, then apply its findings

Reach for the other skills when the job gets more specific:

• craft-autoresearch — deliberate, eval-backed experimentation when you bring test inputs and measurable criteria; explicit-only, never a default next step.
• craft-skill-spec — decide the shape of a new skill-like artifact before writing it.
• craft-harness — plan or repair repo-local agent guidance and related surfaces.
• craft-handoff — end a long session with a durable doc plus a resume prompt.
• spec-charter, spec-system-map, spec-grill — land a brownfield repo spec axis from direction to system shape to capability contracts.

The methodology

The skills are the delivery vehicle; the durable part is two review disciplines that hold up as models get more capable — a smarter agent is exactly what finds the loophole in a loose contract or talks a loop past "good enough." Both are written up as standalone references you can apply without adopting any skill:

• docs/methodology/predicate-test.md — the 3-axis test (Authority / Distributional / Manipulability) for deciding whether a written contract is safe for an agent to optimize against. Applied by spec-grill.
• docs/methodology/loop-stop-conditions.md — falsifiable exit conditions (Self-LGTM / persistent fixpoint / no-op / hard cap) for agent improvement loops. Applied by craft-autoresearch.

Install

CraftKit installs as Agent Skills for Claude Code, and each skill is also a plain SKILL.md file that can be used by Codex and other compatible agents.

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 skills

Skill	Use when	Side effect
craft-prompt	a new prompt is needed from scratch for any LLM or agent interface	returns copy-pasteable text
craft-skill-spec	a new skill needs a concrete spec based on current CraftKit skill-radar judgments before writing SKILL.md	returns a spec; reads radar references
craft-harness	a repo's agent guidance needs placement decisions, cleanup, sync, or a gated change plan across Codex and Claude Code	returns a repo-local plan or small markdown/skill edits; gates risky surfaces
craft-critique	an existing prompt or skill needs a review — read-only by default — before editing or shipping	surfaces strengths, prioritized findings, and fix ordering; applies the fixes only when asked
craft-autoresearch	a prompt or skill works "sometimes" and needs eval-driven iteration	runs evals and may edit mutable files
craft-handoff	a session is ending and the next session needs a copy-paste-ready continuation prompt	writes handoff files and may copy to clipboard
spec-charter	a repo needs a project-wide spec axis for direction, Objectives, Decisions, or stale-spec reassessment	creates or amends spec/charter.md
spec-system-map	a brownfield repo needs high-level system shape, runtime boundaries, flows, invariants, and candidate capability boundaries	creates or amends spec/system-map.md
spec-grill	candidate repo boundaries need to become accepted capability contracts with Behaviors and Hard Constraints	creates or refines spec/capabilities.md after evidence review

When two skills could trigger, choose the least invasive one that answers the request: review-only and apply/fix/improve wording both go to craft-critique (read-only by default; applies fixes when asked); repeated measurable failures go to craft-autoresearch; prior-art questions about skill design go to the craft-skill-spec radar; repo harness placement and Codex/Claude setup work goes to craft-harness.

Terminology note: craft-harness means repo-local agent guidance and provider surfaces. craft-autoresearch uses an eval runner for replaying test inputs and scoring outputs. Do not use "harness" for both.

The spec-* skills form a pipeline: spec-charter -> spec-system-map -> spec-grill. Use them when a brownfield repo needs a compact spec axis grounded in real repo evidence instead of a generic architecture document.

The spec axis supersedes dev-backlog's retired backlog-charter skill (dev-backlog split that surface into the spec-series in its 0.6.0): spec/charter.md is the successor home for the project reference axis, and spec-charter's amend mode reads a legacy root CHARTER.md as a fallback and migrates it deliberately rather than silently. dev-backlog consumes the axis — it measures sprints and triage against spec/charter.md — but does not own it.

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.

Maintainer status

Four skills (craft-prompt, craft-critique, craft-autoresearch, craft-handoff) have been optimized through craft-autoresearch passes against eval suites — including craft-autoresearch itself (reflexive meta-pass). craft-skill-spec, craft-harness, and the spec-* skills are newer and have maintainer-local or repo-local contract evidence, but have not yet been through full autoresearch passes. Publicly reproducible status and local-maintainer evidence boundaries are tracked in docs/status.md.

What belongs in CraftKit

• generating new prompts from scratch (task, research, session handoff, templates)
• prompt design and restructuring
• reusable skill design
• repo spec-axis creation for charter, system map, and capability contracts
• project-specific agent harness design and maintenance
• diagnostic review and minimal-diff editing
• iterative improvement loops
• radar-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 guides or reference files)
5. Eval-driven improvement when possible
6. Copy-pasteable results over fancy abstractions
7. Weight follows durability — as models improve, move each skill's center of gravity from "tell the model how to think" toward "give the model durable state and direction it cannot hold on its own"

Principle 7 is the axis CraftKit is actively re-sized against: machinery (deterministic paths, clipboard, hooks, archiving), time-sensitive curated knowledge (radar, platform surfaces), and direction-setting judgment contracts are model-independent and stay; raw prescription erodes as models improve and gets cut. The 2026-07 right-sizing pass applied it — removing craft-survey, shrinking craft-harness, and loosening craft-critique and craft-handoff toward judgment over fixed format (docs/prd-2026-07-rightsizing.md).

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.

craft-harness also ships reviewable hook asset recipes under skills/craft-harness/assets/hooks/. They provide shared .agents/hooks/scripts/ scripts plus Codex and Claude adapter snippets; they are intentionally not auto-installers.

Skill spine budget

AGENTS.md keeps an absolute 500-line format ceiling for each SKILL.md, but CraftKit's release gate is stricter: npm run verify fails when a skill spine exceeds 220 lines or a frontmatter description exceeds 50 words.

• Normal skills: about 100-160 lines.
• Complex loop or orchestration skills: about 160-220 lines.
• Anything growing past that should move examples, platform notes, maintenance commands, or edge-case catalogs into references/.

The spine should still be understandable alone: purpose, inputs, steps, output contract, one compact example, limitations, and links to on-demand references. References carry depth; the spine carries the operating path. Mirrored references are allowed only when the verifier guards them against drift.

See docs/skill-anatomy.md for the canonical per-family section contract each skill is normalized against.

Invocation policy

Most CraftKit skills are explicit workflow selectors, not always-on background guidance. Use implicit invocation only when a skill is low-risk and broadly helpful when matched, such as read-only diagnosis or direct prompt drafting.

For explicit-only workflows, pair both platform controls:

# SKILL.md frontmatter, used by Claude Code
disable-model-invocation: true

# agents/openai.yaml, used by Codex
policy:
  allow_implicit_invocation: false

Use explicit-only policy for skills that edit files, write artifacts, mutate clipboard state, run eval loops, create spec files, inspect repo harness surfaces, or otherwise turn a broad user request into a higher-ceremony workflow. Keep the description concise and useful for manual skill lists even when it is not injected for implicit routing.

Routing checks

Use these lightweight checks after editing skill descriptions or routing boundaries. They are manual contract checks, not a new runtime.

Prompt	Expected skill	Failure signal
"review this skill, don't edit"	craft-critique	edits the artifact despite the read-only default
"improve this skill and apply changes"	craft-critique (findings, then apply)	stops at read-only findings despite the apply request
"run measured iterations on failures"	craft-autoresearch	describes repo harness setup instead of an eval runner
"set up Codex + Claude repo guidance"	craft-harness	installs or enables hooks/MCP/plugins without an approval gate
"write a prompt for GPT"	craft-prompt	refuses to deliver a copy-pasteable prompt

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. For implicit skills, keep the frontmatter description line as context so the agent knows when to apply the skill.
4. For explicit-only skills, keep the description in the file for menus and manual selection, but preserve the invocation policy fields above when the target agent supports them.

Verify

Run the repo-local smoke check before release or packaging changes:

npm run verify

It checks JSON syntax, package boundaries, skill frontmatter, SKILL.md line budgets, terminology leaks, required README/status paths, and npm pack --dry-run.

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.