AICraft is an open engineering standard and operational discipline for AI coding agents — Antigravity, Claude Code, Cursor, Windsurf, Codex, and others. It turns raw code generation into a disciplined, architecture-respecting engineering process.

What is an AI skill in AICraft?

AICraft installs as a single named skill ("aicraft") built from four pillars — Constitution, Workflow, Playbook, Prompt Library — plus nine deep subguides covering codebase discovery, architecture, task planning, implementation, testing, code review, security, database review, and AI behavior. Each is a structured markdown module the agent reads before acting.

How is AICraft different from a prompt?

A prompt is a temporary, one-off instruction. AICraft is a standing discipline: the same Constitution, workflow, and scenario playbooks apply to every task and every model, producing a repeatable process instead of a one-time answer.

How do I install AICraft?

Run the universal installer (curl ... install.sh | bash) and it auto-detects your environment, or target a specific client by passing its name (claude, cursor, codex, antigravity, windsurf) as an argument. See Install AICraft on this page for exact commands.

Can I use AICraft with Claude Code?

Yes. The installer copies the skill to ~/.claude/skills/aicraft/, and it's invoked with /aicraft.

Can I use AICraft with Codex?

Yes. The installer appends the Constitution, Workflow, Playbook, and subguides to AGENTS.md, which Codex reads automatically.

Can I use AICraft with Cursor?

Yes. The installer generates rule files in .cursor/rules/ that Cursor applies automatically.

規

The AI Forge

AICraft

Craft AI skills instead of simply using AI.

AICraft is built around the idea that AI capabilities should be structured, reusable, composable, and engineered rather than recreated from scratch every time.

Install AICraft View on GitHub

What Is AICraft?

An engineering discipline, installed as a skill.

AICraft is an open engineering standard for AI coding agents — Antigravity, Claude Code, Cursor, Windsurf, Codex, Gemini, and ChatGPT. Its premise: modern models can write syntax easily, but raw generation without discipline produces architectural drift, phantom abstractions, broken domain invariants, and unverified test claims.

In AICraft's own terms, the whole discipline installs as a single named skill — aicraft — built from four composable pillars: a Constitution of non-negotiable rules, a Workflow that every task moves through, a Playbook of scenario-specific runbooks, and a Prompt Library of deterministic templates — plus nine deep subguides for specific domains like architecture, testing, and security.

These pieces are reusable by design: the same Constitution and Workflow apply to every task, and the Playbook composes the right subguides for whichever scenario is in front of the agent, so the discipline doesn't have to be re-explained from scratch each time.

Why AICraft

A prompt is temporary. A skill is repeatable.

One-off prompt

Prompt
↓
Temporary instruction

Answers this one request, in whatever way the model happens to interpret it. Nothing carries over to the next request.

AICraft skill

Skill
↓
Reusable capability
↓
Repeatable workflow
↓
Consistent result

The same Constitution, Workflow, and Playbook apply every time, regardless of which model is running them.

How AICraft Works

Four pillars, read in order. One workflow, every task.

01
Constitution
15 mandatory, non-negotiable engineering rules and a violation policy. Core principle: understand first, build second. If a task conflicts with the Constitution, the Constitution always wins.
02
Workflow
The standard 7-phase execution lifecycle every task moves through: Receive → Understand → Plan → Implement → Validate → Review → Complete.
03
Playbook
8 scenario-specific runbooks — Documentation, Feature Development, Bug Fix, Refactoring, Architecture Modification, Technical Research & Spike, Documentation Review, Code Review — each with its own checklist and explicit "never" list.
04
Prompt Library
Deterministic, reusable prompt schemas and templates, so the same kind of request produces the same structured output regardless of which model is answering it.

Standard 7-Phase Workflow

ReceiveUnderstandPlanImplementValidateReviewComplete

Every task moves through this lifecycle: understand the architecture and domain before planning, implement the smallest correct change, validate with real compiler/linter/ test output, run an 8-stage priority review, then close out with updated docs and an atomic commit.

Skill Architecture

The actual repository structure

The installable skill lives entirely under skill/ — markdown only, nothing that executes:

skill/
├── SKILL.md            # Entrypoint specification
├── constitution.md     # The 15 rules & violation policy
├── workflow.md          # Standard 7-phase execution workflow
├── playbook.md          # 8 scenario runbooks & review priority
├── prompt-library.md    # Deterministic prompt schemas
└── skills/              # Deep specialized subguides
    ├── codebase-discovery.md
    ├── architecture.md
    ├── task-planning.md
    ├── implementation.md
    ├── testing.md
    ├── code-review.md
    ├── security.md
    ├── database-review.md
    └── ai-behavior.md

The 15 Constitution rules

01Read before you write.
02Documentation is the source of truth.
03Tasks drive development.
04Respect the architecture.
05Protect existing decisions.
06Reuse before creating.
07Keep changes atomic.
08Update documentation.
09Think long term.
10Explain decisions.
11Never break the Domain.
12Ask when unsure.
13Respect time.
14Leave the project better.
15Protect the vision.

Nine deep subguides

codebase-discovery.md
architecture.md
task-planning.md
implementation.md
testing.md
code-review.md
security.md
database-review.md
ai-behavior.md

The Playbook composes the relevant subguide(s) for each scenario — a Code Review pulls in code-review.md and, where relevant, security.md — rather than the agent working from the Constitution alone.

Ground Truth Non-Negotiables

The agent must never claim: tests passed when they weren't executed, an integration works when it wasn't verified with real output, a requirement exists when it wasn't specified, a file changed when it was untouched, an architectural pattern was approved when no ADR exists, or an error was fixed when it was merely suppressed in a try/catch.

AICraft does not currently document an end-user workflow for authoring brand-new custom skills from a template — its documented skill surface is the AICraft discipline above, installed as-is. This section will be expanded here if the repository adds that capability.

Installation

Install AICraft

One universal installer auto-detects your environment (Claude Code, Cursor, Codex, Antigravity, Windsurf) and installs into all of the ones it finds:

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash

Or target one client directly by passing its name as an argument — the commands below do exactly that.

Per-client installation

Claude Code

Supported

Installation

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash -s -- claude

Configuration — installs to ~/.claude/skills/aicraft/.

Usage — Invoke with /aicraft once installed.

Verification — Run /aicraft in Claude Code and confirm it responds with the AICraft discipline instead of an unknown-command error.

Cursor

Supported

Installation

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash -s -- cursor

Configuration — installs to .cursor/rules/aicraft.md.

Usage — Cursor applies the rule automatically once it's in .cursor/rules/.

Verification — Check that .cursor/rules/ now contains the AICraft constitution, workflow, and playbook files.

Codex

Supported

Installation

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash -s -- codex

Configuration — installs to AGENTS.md.

Usage — Codex reads the appended AICraft block from AGENTS.md automatically.

Verification — Open AGENTS.md and confirm an "AICraft Discipline" section was appended.

Antigravity

Supported

Installation

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash -s -- antigravity

Configuration — installs to .agents/skills/aicraft/ (workspace) and ~/.gemini/config/skills/aicraft/ (global).

Usage — Antigravity picks up the installed skill from either location.

Verification — Confirm the aicraft/ folder now exists under .agents/skills/ or ~/.gemini/config/skills/.

The same installer also supports Windsurf (-s -- windsurf), writing directly to .windsurfrules.

Quick Start

Fastest path to a disciplined agent

Run the universal installer from your project root:

curl -fsSL https://raw.githubusercontent.com/bishoy-bishai/AICraft/main/install.sh | bash

02
In Claude Code, run /aicraft. In Cursor, Codex, or Antigravity, just start working — the installed rule/instruction file is read automatically.
03
Give it a real task ("fix this bug," "add this feature," "review this branch"). The agent runs the Pre-Implementation Gate first — problem, architecture, domain, task — before writing anything.

How The Pillars Compose

Not one prompt — four layers working together

A single task draws on more than one pillar at once. A code review, for example, combines:

Constitution

Workflow · Review phase

Playbook · Scenario 8

code-review.md + security.md

One consistent, evidence-grounded review

Compatibility

Supported AI Coding Agents

Claude CodeSupported
CursorSupported
CodexSupported
AntigravitySupported

All four have a dedicated installer target in the repository. Windsurf is also supported through the same installer.

Examples

Why a skill beats a one-off prompt

User request

"Add a discount field to the checkout flow."

Without AICraft

A raw prompt gets code back immediately — but with no guarantee it respected the existing layering, reused the right primitives, or came with a test.

AICraft approach

AICraft's Feature Development scenario (Playbook #2) forces the Pre-Implementation Gate first: does the agent understand the problem, the architecture, the domain, the task? Only then does it implement — atomically, with tests, matching existing patterns.

Expected outcome

A bounded, reviewed change with unit/integration coverage and updated docs — not just code that compiles.

User request

"Fix the bug where totals round incorrectly."

Without AICraft

A raw prompt might patch the symptom without isolating the actual root cause.

AICraft approach

The Bug Fix scenario (Playbook #3) requires isolating the root cause and writing a failing reproduction test before applying the minimal correct fix.

Expected outcome

A fix with a regression test attached, verified against the existing suite — not a guess that happens to look right.

User request

"Review this pull request."

Without AICraft

An unstructured review can drift toward style nitpicks and miss real risk.

AICraft approach

The Workflow's Review phase runs an 8-stage priority order — Architecture → Domain → Correctness → Security → Performance → Readability → Testing → Docs — pulling in the code-review.md and security.md subguides as needed.

Expected outcome

A review that catches architectural and security issues before style issues, in a consistent order every time.

User request

"Refactor this module."

Without AICraft

A raw prompt can't tell refactoring apart from feature work, and may mix the two.

AICraft approach

The Refactoring scenario (Playbook #4) keeps refactors isolated from feature changes — one of the Constitution's Ground Truth rules the agent is never allowed to blur.

Expected outcome

A clean, reviewable refactor diff with no unrelated feature work riding along inside it.

FAQ

Frequently asked questions

What is AICraft?: AICraft is an open engineering standard and operational discipline for AI coding agents — Antigravity, Claude Code, Cursor, Windsurf, Codex, and others. It turns raw code generation into a disciplined, architecture-respecting engineering process.
What is an AI skill in AICraft?: AICraft installs as a single named skill ("aicraft") built from four pillars — Constitution, Workflow, Playbook, Prompt Library — plus nine deep subguides covering codebase discovery, architecture, task planning, implementation, testing, code review, security, database review, and AI behavior. Each is a structured markdown module the agent reads before acting.
How is AICraft different from a prompt?: A prompt is a temporary, one-off instruction. AICraft is a standing discipline: the same Constitution, workflow, and scenario playbooks apply to every task and every model, producing a repeatable process instead of a one-time answer.
How do I install AICraft?: Run the universal installer (curl ... install.sh | bash) and it auto-detects your environment, or target a specific client by passing its name (claude, cursor, codex, antigravity, windsurf) as an argument. See Install AICraft on this page for exact commands.
Can I use AICraft with Claude Code?: Yes. The installer copies the skill to ~/.claude/skills/aicraft/, and it's invoked with /aicraft.
Can I use AICraft with Codex?: Yes. The installer appends the Constitution, Workflow, Playbook, and subguides to AGENTS.md, which Codex reads automatically.
Can I use AICraft with Cursor?: Yes. The installer generates rule files in .cursor/rules/ that Cursor applies automatically.
How do I create an AI skill with AICraft?: AICraft does not currently document an end-user workflow for authoring new custom skills from scratch. Its documented skill surface is the AICraft discipline itself — the four pillars and nine subguides described above — which you install as-is into your agent of choice.

The AI Forge

AICraft

Craft reusable intelligence.

Install AICraft View on GitHub

Part of the SIMURAI Arsenal