Skip to content

nolrm/contextkit

Repository files navigation

ContextKit

Context Engineering for AI Development

Give your AI assistants (Cursor, Claude, Copilot, Codex, OpenCode, Gemini, Aider, Continue, Windsurf) structured context through markdown files. ContextKit creates a knowledge base that ensures AI generates code matching your exact patterns, style, and architecture—no more hallucinated code or mismatched conventions.

ContextKit is a CLI tool that provides context-engineering capabilities by creating .contextkit/ directories with project standards, guidelines, and patterns that AI assistants read automatically.

Read the full documentation · How context works

Why ContextKit?

The problem: LLMs are great at syntax, not at your conventions. Generic AI output requires manual fixes for style, structure, and architecture.

The solution: ContextKit provides your AI with:

  • Glossary of project terminology and domain-specific terms (e.g., your entity, feature, and module names)
  • Standards for code style, testing patterns, and architecture
  • Templates with canonical component shapes

Update .md files as your project evolves; the AI follows.

Multi-Platform Support

Works with: CursorClaude CodeGitHub CopilotCodex CLIOpenCodeGemini CLIAiderContinueWindsurf

Each platform gets auto-loaded bridge files (CLAUDE.md, AGENTS.md, GEMINI.md, .windsurfrules, etc.) so your AI tools read project standards automatically. Claude Code uses @ imports in CLAUDE.md to load standards content directly into context — no extra token cost from manual file reads.


Quick Start (60s)

1. Install the CLI

npm i -g @nolrm/contextkit

2. Set up your project

cd your-project
contextkit install

Creates .contextkit/ with skeleton standards files, a self-describing README.md, and an attribution block in config.yml so any developer who encounters the folder knows what manages it.

3. Generate your standards

Run /analyze in your AI tool — it scans your codebase and fills the skeleton files with your project's conventions.

Done. Your AI tools now have project-specific context.


Multi-Team Workflow

Perfect for teams where members use different AI tools:

# First team member - sets up the project with their tool
contextkit install claude   # or: contextkit install (interactive picker)

# Each additional team member adds their platform
ck claude      # creates CLAUDE.md + .claude/rules/
ck cursor      # creates .cursor/rules/ (scoped .mdc files)
ck copilot     # creates .github/copilot-instructions.md
ck codex       # creates AGENTS.md
ck opencode    # creates AGENTS.md
ck gemini      # creates GEMINI.md + .gemini/settings.json
ck aider       # creates CONVENTIONS.md + .aider/rules.md
ck continue    # creates .continue/rules/ + config.yaml
ck windsurf    # creates .windsurfrules + .windsurf/rules/
ck vscode      # alias for copilot

Each platform generates bridge files that the AI tool auto-reads. If a bridge file already exists (e.g., you have a custom CLAUDE.md), ContextKit appends its section below your content instead of overwriting. Share your .contextkit/standards/*.md files with the team and everyone gets the same context.


See the difference (before → after)

Prompt

"Add checkout flow for customer"

What the AI does with ContextKit

  • Reads glossary.mdcheckout = checkout process; customer = customer account
  • Applies code-style.md → strict TS, functional components
  • Follows testing.md → numbered test cases

Result (diff)

- const Checkout = () => <button>Buy</button>
+ export function CheckoutFlow({ customer }: { customer: string }) {
+   // Uses customer from glossary context
+   return <div>Checkout for {customer}</div>
+ }

Use it in your tool

Cursor — rules auto-load from .cursor/rules/, slash commands in .cursor/prompts/

/analyze    # scan codebase and generate standards
/review     # code review with checklist
/fix        # diagnose and fix bugs

Claude CodeCLAUDE.md uses @ imports to auto-load all standards into context every session (no manual reads needed, saves tokens). Slash commands in .claude/commands/.

/analyze    # scan codebase and generate standards
/review     # code review with checklist
claude "create checkout flow for customer"

GitHub Copilot — reads .github/copilot-instructions.md automatically

@.contextkit Create checkout flow for customer

Codex CLI — reads AGENTS.md automatically

codex "create checkout flow for customer"

OpenCode — reads AGENTS.md automatically

opencode "create checkout flow for customer"

Slash Commands

ContextKit installs reusable slash commands for supported platforms:

Command What it does
/analyze Scan codebase and generate standards content
/review Code review with checklist
/fix Diagnose and fix bugs
/refactor Refactor code with safety checks
/test Generate comprehensive tests
/doc Add documentation
/spec Write a component spec (MD-first) before any code is created
/squad Kick off a squad task — one task or many (auto-detects batch mode). Pushes back with clarifying questions if the task is vague.
/squad-architect Design the technical plan from the PO spec
/squad-dev Implement code following the architect plan
/squad-test Write and run tests against acceptance criteria
/squad-review Review the full pipeline and give a verdict
/squad-doc Create companion .md files for new/modified code after review passes
/squad-auto Auto-run the full pipeline after kickoff (recommended, sequential)
/squad-auto-parallel Auto-run the pipeline in parallel using Claude Code agents (Claude Code only)
/ck Health check — verify setup, standards, and integrations

Claude Code — available as /analyze, /review, etc. in .claude/commands/ Cursor — available as slash commands in Chat via .cursor/prompts/

Both platforms delegate to the universal command files in .contextkit/commands/, so you maintain one set of workflows.


Squad Workflow

The squad workflow turns a single AI session into a structured multi-role pipeline. Each role has its own slash command that reads and writes to a shared handoff file (.contextkit/squad/handoff.md), simulating a team of specialists.

Pipeline Roles

Step Role Command What it does
1 Product Owner /squad Writes a user story, acceptance criteria, edge cases, and scope. If the task is ambiguous, asks up to 5 clarifying questions before writing the spec. Optionally captures screenshots/images as visual assets.
2 Architect /squad-architect Designs the technical approach, files to change, and implementation steps
3 Developer /squad-dev Implements the code following the architect's plan
4 Tester /squad-test Writes and runs tests against the PO's acceptance criteria
5 Reviewer /squad-review Reviews everything and gives a PASS or NEEDS-WORK verdict
6 Doc Writer /squad-doc Creates companion .md files for every new/modified code file

Single-Task Flow

/squad "add dark mode support"   # PO writes the spec

/squad-auto                      # Auto-runs architect → dev → test → review → doc (recommended)
# — or step through manually —
/squad-architect                 # Architect designs the plan
/squad-dev                       # Dev implements the code
/squad-test                      # Tester writes and runs tests
/squad-review                    # Reviewer gives the verdict
/squad-doc                       # Doc Writer creates companion .md files

Batch Flow

Pass multiple tasks to /squad and it automatically runs in batch mode:

/squad "add dark mode" "fix login bug" "refactor checkout"
# PO writes specs for all three tasks

/squad-auto
# Runs Architect → Dev → Test → Review → Doc for each task sequentially

Parallel mode (Claude Code only): Use /squad-auto-parallel instead of /squad-auto to spawn parallel subagents — one per task per phase — so all tasks progress simultaneously rather than one at a time.

/squad "add dark mode" "fix login bug" "refactor checkout"

/squad-auto-parallel
# Phase 1: architect agents for all 3 tasks run in parallel
# Phase 2: dev→test→review pipeline runs in parallel per task; Phase 3: doc runs sequentially

Model routing (Claude Code only): Set model_routing: true in .contextkit/squad/config.md to have /squad-auto automatically use Claude Haiku for Dev and Test phases. Architect and Review always run on your primary model. Saves ~35% tokens with no quality loss — the standards files and Review gate maintain quality.

# .contextkit/squad/config.md
checkpoint: po
model_routing: true   # dev + test → Haiku, architect + review → primary model

Feedback Loop

Any downstream role can raise questions for an upstream role. When this happens, the pipeline pauses and directs you to the right command:

Reviewer has questions for Dev → run /squad-dev to clarify
Tester has questions for Architect → run /squad-architect to clarify
Architect has questions for PO → run /squad to clarify

After clarifications are added, re-run the asking role's command to continue. This prevents misunderstandings from compounding through the pipeline.

Visual Assets (Optional)

If you have a screenshot, mockup, or design image relevant to the task, paste or attach it when running /squad. The PO agent will save it to .contextkit/squad/assets/ and reference the path in the handoff. Architect and Dev agents automatically read any listed assets when they pick up the handoff.


Git Hooks & Quality Gates

ContextKit can optionally install Git hooks during ck install. Uses git config core.hooksPath to point Git at .contextkit/hooks/ — no external dependencies like Husky required. Works in any git repo, not just Node.js projects.

For Node.js projects, a prepare script is automatically added to package.json so hooks activate for all developers after npm install — no need for everyone to run ck install.

Hook What it does
pre-push Quality Gates — auto-detects your project framework and runs the appropriate checks
commit-msg Enforces Conventional Commits format

Framework-Aware Quality Gates

The pre-push hook detects your project type and runs the right quality checks automatically. All gates skip gracefully when tools aren't installed.

Framework Checks
Node.js TypeScript, ESLint, Prettier, build, test (auto-detects npm/yarn/pnpm/bun)
Python ruff/flake8, mypy, black/ruff format, pytest
Rust cargo check, clippy, cargo test
Go go vet, golangci-lint, go test
PHP PHPStan, PHPUnit
Ruby RuboCop, RSpec/rake test
Java Maven verify / Gradle check

Commit Message Format

When the commit-msg hook is enabled, all commits must follow this format:

<type>(<scope>): <description>

Types: feat, fix, improve, docs, style, refactor, test, chore

Examples:

git commit -m "feat(auth): add login page"
git commit -m "fix: resolve null pointer in checkout"
git commit -m "docs: update API reference"
git commit -m "test(cart): add edge case coverage"

Hooks are optional and can be skipped with ck install --no-hooks.


Key Features

  • 🧠 Context Engineering - Structured MD files your AI reads automatically
  • 🔍 Smart Analysis - AI generates standards content based on your codebase
  • 🌍 Project Agnostic - Works with React, Vue, Node.js, PHP, Python, Rust, monorepos—any project type
  • 🤖 Multi-Platform - Works with Cursor, Claude Code, Copilot, Codex, OpenCode, Gemini, Aider, Continue, Windsurf
  • 🛡️ Safe Install - Backs up existing files with rollback support
  • Zero Config - Auto-detects project type and package manager
  • Policy Enforcement - Configurable validation with ck check
  • 📝 Corrections Tracking - Track AI performance issues with corrections log

Commands

# Installation & Setup
ck install            # set up .contextkit + pick AI tool interactively
ck install claude     # set up .contextkit + Claude, or add Claude to an existing install
ck claude      # add Claude Code integration (CLAUDE.md + rules)
ck cursor      # add Cursor integration (scoped .mdc rules)
ck copilot     # add GitHub Copilot integration
ck codex       # add Codex CLI integration (AGENTS.md)
ck opencode    # add OpenCode integration (AGENTS.md)
ck gemini      # add Gemini CLI integration (GEMINI.md)
ck aider       # add Aider integration (CONVENTIONS.md)
ck continue    # add Continue integration
ck windsurf    # add Windsurf integration (.windsurfrules)
ck vscode      # alias for copilot

# Analysis & Updates
/analyze       # customize standards to your project (slash command in your AI tool)
ck update      # pull latest commands/hooks — preserves your analyzed standards
               # updates are also flagged automatically after each ck command (24h cache)
ck status      # check install & integrations

# Validation & Compliance
ck check       # validate installation & policy compliance
ck check --strict  # treat warnings as errors

# Corrections Logging
ck note "message"  # add note to corrections log
ck note "AI issue" --category "AI Behavior" --priority HIGH

# Squad — multi-role AI pipeline (slash commands in your AI tool)
/squad "add dark mode"   # PO writes spec
/squad-auto              # runs architect → dev → test → review → doc hands-free
/squad-reset             # clear stuck or mixed squad state

Links

• 🐛 Issues • 💬 Discussions


License

MIT

Author

Marlon Maniti
GitHub: @nolrm

About

Vibe Kit is a comprehensive development toolkit that provides AI guidance, type safety, quality checks, and automated workflows for modern development teams.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors