| name | scaffold-project |
| description | Bootstrap or review project-level context primitives for Claude Code. First run creates .claude/primitives/ and .claude/skills/project-context/. Re-runs review existing primitives against the codebase, auto-update what changed, and flag what's stale. Invoke with '/scaffold-project' or say 'scaffold project', 'set up project context', 'review project context'. |
Scaffold Project
Dual purpose. First run: creates the primitive files and session-start skill. Re-runs: review existing primitives against the current codebase — auto-updates what changed, flags what's stale, asks about what it can't determine. After scaffolding, /project-context runs automatically at session start via the project AGENTS.md.
Phase 0: Pre-check
-
Confirm cwd looks like a project root. Check for package.json, pyproject.toml, go.mod, Gemfile, or any framework config file. If none found, ask: "This doesn't look like a project root. Are you in the right directory?"
-
Check if .claude/primitives/ already exists. If it does, this is a review run, not a fresh bootstrap. Inform the user:
Primitives already exist — this will be a review pass. Auto-updatable fields (versions, scripts, folder structure) will be refreshed. Manually-curated sections (deployment, gotchas, glossary, architecture) will be flagged if they look stale but not overwritten. Proceed?
-
Determine project type:
package.json exists → Node project. Proceed to Phase 1 (auto-discovery).- No
package.json but other indicators exist (pyproject.toml, go.mod, Gemfile) → Non-Node project. Note the language. Ask the user to confirm stack details (auto-population is limited for non-Node). Proceed to Phase 2 with user-provided info.
- Nothing at all → New/empty project. Ask the key questions (see New Project Questions below). Proceed to Phase 2 with user answers.
Phase 1: Discovery (Node projects — read everything, write nothing)
Read these files. Skip gracefully if missing — don't error, just leave that field blank.
Config files
package.json — full contents (deps, devDeps, scripts, name, engines).env.example or env.example — environment variables.nvmrc — Node versiontsconfig.json or jsconfig.json — path aliases, compiler optionsvite.config.ts / vite.config.js — Vite config (check server.port)next.config.js / next.config.ts — Next.js confignuxt.config.ts — Nuxt configsupabase/config.toml — Supabase project_idtailwind.config.js / tailwind.config.ts — Tailwind config.prettierrc / .prettierrc.js / .prettierrc.json — Prettier configeslint.config.js / .eslintrc / .eslintrc.json / .eslintrc.js — ESLint configREADME.md — first 50 lines only
Lock files (package manager detection)
Check which exists (first match wins): pnpm-lock.yaml → pnpm, yarn.lock → yarn, package-lock.json → npm
Existing context
CLAUDE.md at project root — extract any GitHub config, conventions, active context, deploy info that should migrate to primitives.claude/ directory contents — note what already exists (settings.json, commands/, etc.) — don't overwrite non-primitive files.claude/AGENTS.md — if it exists, read the full file. Extract: any orient table rows not in the standard template, any existing conventions, any project-specific instructions. These will be merged into the new AGENTS.md (see Phase 2)..claude/skills/ — list all subdirectories. Any existing project skill (a directory with a SKILL.md inside) that is NOT project-context is a project-specific skill that must be preserved. Record the skill name and its first-line description (from its SKILL.md frontmatter description field). These will appear in the new AGENTS.md orient table and primitive map..claude/primitives/ — if it exists, read ALL primitive files. This is the baseline for the review diff in Phase 2. Note which optional primitives (glossary.md, architecture.md) already exist.
Optional primitive signals
Two primitives are optional — only create them if signals are present OR the user opts in.
glossary.md — create if ANY of these are true:
- README.md contains a terminology, glossary, or definitions section
- The project domain is specialized (legal, medical, finance, real estate, etc.) — infer from README or package name
- Source files use domain-specific terms that aren't standard programming concepts
- User explicitly opts in
architecture.md — create if ANY of these are true:
- Multiple top-level source directories beyond a single
src/ (e.g., api/, workers/, packages/)
- README has an architecture, overview, or system design section
- Project has multiple services or a monorepo structure
- User explicitly opts in
If signals are detected, note them — they'll be used in Phase 2. If borderline, ask: "This looks like it might benefit from a [glossary / architecture] doc. Want me to create one?"
Convention scan (read 3-5 source files)
- List
src/ directory structure. Note: feature-based (src/features/) vs layer-based (src/components/ + src/pages/)
- Read one
.tsx or .vue component file. Note: import order, export style (default vs named), props pattern
- Check for co-located test files (
.test.ts next to source) vs separate test directory
Derive from all of the above
Use this mapping. If you can't determine a value, leave it as <not detected>.
| Field | How to derive |
|---|
| Node version | .nvmrc content, or package.json engines.node |
| Package manager | Lock file detection above |
| Framework | react in deps = React. next = Next.js. vue = Vue. nuxt = Nuxt. svelte = Svelte |
| Bundler | vite.config.* exists = Vite. next in deps = Next (built-in). webpack.config.* = Webpack |
| TypeScript version | package.json devDeps typescript version string |
| CSS | tailwindcss in deps or tailwind.config.* = Tailwind. styled-components in deps. *.module.css files in src = CSS Modules |
| State management | @tanstack/react-query = TanStack Query. zustand = Zustand. redux or @reduxjs/toolkit = Redux |
| Routing | react-router-dom version. Or next = Next App/Pages Router |
| Supabase refs | supabase/config.toml project_id. Or .env.example SUPABASE_* patterns |
| Test runner | vitest = Vitest. jest = Jest. @playwright/test = Playwright (E2E) |
| Dev port | Vite default 5173. Next default 3000. Check vite.config server.port for override |
| Env vars | .env.example contents. Redact any actual secret values |
| Scripts | All entries from package.json scripts — note what each one does based on name and command |
| Folder structure | src/features/ = feature-based. + = layer-based |
New Project Questions
If the project is new/empty, ask these questions:
- What framework? (React / Next.js / Vue / Nuxt / other)
- Supabase? (yes / no)
- Test runner? (vitest / jest / none)
- TypeScript? (yes / no)
- CSS approach? (Tailwind / CSS Modules / other)
Use answers to populate primitives. Leave sections that don't apply as <not applicable>.
Phase 2: Write or Review Primitives
Two paths. Pick the right one.
If this is a FRESH scaffold (no existing primitives)
Write all files from scratch using discovery data or user answers. Templates and structure are below.
If this is a REVIEW run (primitives already exist)
Don't rewrite from scratch. For each primitive, diff discovered state against the existing file. Apply this logic:
| Field type | Action |
|---|
| Auto-updatable — derived directly from codebase | Update silently. These are: versions, package manager, framework, bundler, scripts, dev port, folder structure, component patterns. |
| Manually-curated — written by hand, can't be re-derived | Flag if stale, don't overwrite. These are: Deployment, Environment Variables, Common Gotchas, MCP Servers, Key Dependencies, Glossary terms, Architecture descriptions. |
| Never touch on review | active-context.md — this is the live session scratchpad. |
After updating, present a staleness report — list every section that looks outdated with a one-line note on why. The user can then /remember updates or edit manually.
The file templates and structure follow. Use them for fresh scaffolds. For review runs, use them as the reference for what each file should look like.
Create .claude/ directory structure
mkdir -p .claude/primitives
mkdir -p .claude/skills/project-context
mkdir -p .claude/rules
mkdir -p .cursor/rules
Don't touch anything else in .claude/ that already exists.
Create workflow enforcement rules
These ensure the workflow manager triggers before the agent responds. Create both files — they target Claude Code and Cursor respectively.
.claude/rules/workflow-manager.md:
# Workflow Manager Protocol
BEFORE responding to any work request, you MUST:
1. Announce: "🎯 Workflow Manager active. Checking project state..."
2. Read and follow `~/.claude/skills/workflow-router/SKILL.md`
3. Use the Quick Start checklist to determine current state
4. Report findings and propose the next skill
5. Wait for human confirmation
Do NOT skip this. Do NOT answer the question first. Orient, then act.
.cursor/rules/workflow-manager.mdc:
---
description: Workflow Manager Protocol - orient before any work
alwaysApply: true
---
BEFORE responding to any work request, you MUST:
1. Announce: "🎯 Workflow Manager active. Checking project state..."
2. Read and follow `~/.claude/skills/workflow-router/SKILL.md`
3. Use the Quick Start checklist to determine current state
4. Report findings and propose the next skill
5. Wait for human confirmation
Do NOT skip this. Do NOT answer the question first. Orient, then act.
Write .claude/AGENTS.md (merge-aware)
This is the only file scaffold-project actively merges rather than overwrites. If an existing .claude/AGENTS.md was read in Phase 1, preserve its unique content. If not, write the template clean.
Merge rules:
- The session-start blockquote, "Orient Here" heading, and "Conventions" section always come from the template below — these are the prescribed structure. Do not keep old versions of these.
- The orient table starts with the 6 standard rows (primitives +
/remember + global skills). Then append any rows from the old orient table that are not duplicates — these are project-specific orient entries the team added manually.
- The Primitive Map always shows the 4 standard primitives. Then append any existing project skills discovered in Phase 1 (name + one-line description from their SKILL.md frontmatter).
- If the old AGENTS.md had any sections beyond Orient Here, Conventions, and Primitive Map (e.g. project-specific notes, team conventions, runbooks), preserve them as-is below the Primitive Map.
Template (fill in, then apply merge rules above):
# <Project Name> — Agent Context
> **MANDATORY FIRST ACTION:** Before responding to ANY work request, read `~/.claude/skills/workflow-router/SKILL.md` and follow its Quick Start checklist. Announce `🎯 Workflow Manager active. Checking project state...`, orient, then propose the next skill. Do NOT skip this step.
Project-level entry point for Claude Code.
---
## Orient Here
| If you need to... | Read this |
|---|---|
| Know the tech stack and versions | `.claude/primitives/stack.md` |
| Know coding conventions for this project | `.claude/primitives/conventions.md` |
| Know how to run dev locally | `.claude/primitives/local-dev.md` |
| Check what's currently in progress | `.claude/primitives/active-context.md` |
| Store a fact or instruction | Use `/remember` |
| Use global workflow skills | `~/.claude/skills/AGENTS.md` |
<!-- OPTIONAL: add if glossary.md was created -->
<!-- | Look up domain terms or acronyms | `.claude/primitives/glossary.md` | -->
<!-- OPTIONAL: add if architecture.md was created -->
<!-- | Understand system architecture and data flow | `.claude/primitives/architecture.md` | -->
<!-- MERGE: append non-duplicate rows from old orient table here -->
---
## Conventions
1. **Primitives are factual snapshots.** They describe what the project IS. Update them when the project fundamentally changes, or use `/remember` for one-off facts.
2. **active-context.md is the scratchpad.** The only primitive that changes every session. Update it as you work.
3. **Global skills handle workflow.** `/interview`, `/implement-direct`, `/diagnose`, `/qa-handoff` etc. live at `~/.claude/skills/`. Don't re-implement them here.
4. **Other primitives are on-demand.** Don't load stack.md, conventions.md, or local-dev.md at session start. Read them when you actually need them (e.g., before writing code, before running dev).
---
primitives/
stack.md ← frameworks, versions, key dependencies, MCP servers
conventions.md ← file naming, component patterns, import order
local-dev.md ← commands, ports, env vars, scripts, deployment
active-context.md ← current branch, active issues, decisions [MUTABLE — update every session]
glossary.md ← domain terms, acronyms, jargon
architecture.md ← system structure, data flow, module relationships
<!-- MERGE: append existing project skills here, e.g.:
skills/
my-skill/ ← <description from SKILL.md frontmatter>
-->
<!-- MERGE: append any extra sections from old AGENTS.md here -->
Write .claude/primitives/stack.md
# Tech Stack
**Last Updated:** <today's date YYYY-MM-DD>
---
## Runtime
- **Node:** <version>
- **Package Manager:** <npm | yarn | pnpm>
## Frontend
- **Framework:** <React <version> | Next.js <version> | Vue <version> | ...>
- **Bundler:** <Vite <version> | Webpack | Next.js built-in>
- **Language:** <TypeScript <version> | JavaScript>
- **UI Components:** <shadcn/ui | Radix | none | ...>
- **CSS:** <Tailwind CSS <version> | CSS Modules | styled-components | ...>
- **State:** <TanStack Query <version> | Zustand <version> | Redux | none>
| Next.js App Router | Next.js Pages Router | ...>
Dev project ref:
<!-- Add MCP servers and tools here. Use /remember to update. -->
<!-- Non-obvious dependencies an agent might not know about. Use /remember to add. -->
Write .claude/primitives/conventions.md
# Coding Conventions
**Last Updated:** <today's date YYYY-MM-DD>
---
## File and Folder Structure
- **Organization:** <Feature-based (`src/features/<name>/`) | Layer-based (`src/components/`, `src/pages/`)>
- **Component files:** <PascalCase (`MyComponent.tsx`) | camelCase>
- **Utility files:** <camelCase (`formatDate.ts`) | ...>
- **Test files:** <co-located with `.test.ts` suffix | in `__tests__/` subdirs | in `test/` at root>
## Imports
- **Order:**
1. Node/external packages
2. <Path-aliased internal modules (e.g., `@/components/...`), if aliases exist>
3. Relative imports
Conventional commits: , , , etc.
>
>
Write .claude/primitives/local-dev.md
# Local Development
**Last Updated:** <today's date YYYY-MM-DD>
---
## Quick Start
```bash
# 1. Install dependencies
<npm install | yarn | pnpm install>
# 2. Set up environment
cp .env.example .env.local
# Edit .env.local with actual values (see Environment Variables below)
# 3. Start backend (if applicable)
<supabase start — or remove this step if not applicable>
# 4. Start dev server
<npm run dev | yarn dev | pnpm dev>
Dev Server
Backend / Database
Environment Variables
Required for local dev. See .env.example for full list.
Scripts
What each package.json script does:
| Script | Purpose |
|---|
<script name> | |
Testing
| What | Command |
|---|
| Unit tests | <from package.json scripts.test> |
| Single test | <e.g., npx vitest <pattern>> |
| <E2E tests — if applicable> | <command> |
Deployment
Build
<npm run build | yarn build | pnpm build>
Common Gotchas
### Write `.claude/primitives/active-context.md`
```markdown
# Active Context
> Updated every session. Single source of truth for what is happening right now.
---
## Current State
- **Branch:** <current git branch>
- **Sprint / Focus:** <not set>
## Active Issues
| # | Title | Status | Notes |
|---|-------|--------|-------|
| | | | |
## Recent Decisions
| Date | Decision | Rationale |
|------|----------|-----------|
## Blockers
- <none>
---
> **Usage:** This is read automatically at session start (see AGENTS.md). Update Active Issues and Blockers as you work. Log decisions in Recent Decisions. Use `/remember` to add facts to other primitives.
Write .claude/primitives/glossary.md (optional)
Only create if detection signals were found in Phase 1 or user opted in. Seed with terms discovered during convention scan and README review — don't leave it empty, an empty glossary is worse than no glossary.
# Glossary
> Domain terms, acronyms, and jargon for this project. If a term could be misunderstood or conflicts with common usage, it belongs here. Use `/remember` to add terms as you encounter them.
**Last Updated:** <today's date YYYY-MM-DD>
---
| Term | Means | Notes |
|------|-------|-------|
| <term> | <plain-English definition> | <where it appears, or common misinterpretation to avoid> |
Write .claude/primitives/architecture.md (optional)
Only create if detection signals were found in Phase 1 or user opted in. Focus on the structure that an agent wouldn't be able to infer from just reading files — how do the pieces connect? What's the data flow? Where are the boundaries?
# Architecture
> How this system is structured and how the pieces connect. Not what frameworks we use (that's stack.md) — this is WHY things are organized the way they are.
**Last Updated:** <today's date YYYY-MM-DD>
---
## System Overview
<1-3 sentence summary of what the system does and its main components>
## Module Map
<Describe the main modules/directories and what each one owns. Use a simple list or diagram — not a full file tree.>
## Data Flow
<How data moves through the system. Key: where do user inputs enter? Where does data get stored? What triggers what? Focus on the non-obvious paths.>
## Key Boundaries
Write .claude/skills/project-context/SKILL.md
---
name: project-context
description: "Full project orientation. Reads active context, checks git state, and presents a summary. Also runs automatically via AGENTS.md at session start — invoke explicitly with '/project-context' if you need a fresh orientation mid-session or want to update active-context."
allowed-tools: Read, Grep, Glob, Bash
---
# Project Context
Full orientation for this project. Runs automatically at session start via AGENTS.md. Invoke explicitly with `/project-context` if you need a fresh read mid-session.
## Instructions
### Step 1: Read Active Context
Read `.claude/primitives/active-context.md`.
### Step 2: Check Git State
Run:
```bash
git status
git branch --show-current
git log --oneline -5
Step 3: Update Active Context
If the current branch differs from what's recorded, update the Branch field in active-context.md.
Step 4: Present Summary
Output this structure:
Project:
Branch:
Active Issues:
Recent Decisions:
<last 3 from active-context, or "None recorded">
Blockers:
<from active-context, or "None">
Need more context?
- Stack & versions →
.claude/primitives/stack.md
- Coding conventions →
.claude/primitives/conventions.md
- Dev commands & scripts →
.claude/primitives/local-dev.md
Ready. What would you like to work on?
What This Skill Does NOT Do
- Does not load stack.md, conventions.md, or local-dev.md (those are on-demand)
- Does not start dev servers
- Does not pull from git
Phase 3: Confirm
Print a summary appropriate to whether this was a fresh scaffold or a review run.
Fresh scaffold output:
Scaffold complete.
Created:
.claude/rules/workflow-manager.md ← Claude Code workflow enforcement
.cursor/rules/workflow-manager.mdc ← Cursor workflow enforcement
.claude/AGENTS.md ← project entry point
.claude/primitives/stack.md ← <one-line stack summary>
.claude/primitives/conventions.md ← <org pattern, key conventions>
.claude/primitives/local-dev.md ← <dev command, port>
.claude/primitives/active-context.md ← empty, fill as you work
.claude/skills/project-context/SKILL.md ← explicit orientation skill
<if created> .claude/primitives/glossary.md ← <N terms seeded>
<if created> .claude/primitives/architecture.md ← <one-line summary>
Session start is automatic — AGENTS.md instructs the agent to orient on first load.
Use /remember to add facts as you discover them.
Review run output:
Review complete.
Auto-updated:
<list each field that was refreshed and what changed, e.g. "stack.md: TypeScript 5.1 → 5.3">
(or "Nothing changed — primitives are current." if nothing moved)
⚠ Stale (needs your attention):
<list each section that looks outdated with a one-line note, e.g. "local-dev.md Deployment: last updated 2025-11-02, no deploy commands found in package.json">
(or "Nothing flagged." if everything looks fresh)
Preserved (merged into AGENTS.md):
<list any existing project skills and custom orient rows carried over>
(or "Nothing to merge." if clean)
Use /remember to update any flagged sections.
What This Skill Does NOT Do
- Does not modify or delete existing
CLAUDE.md
- Does not touch existing
.claude/ files other than AGENTS.md (settings.json, commands/, rules/, existing project skills, etc. are all preserved)
- Does not delete or overwrite existing project skills in
.claude/skills/ — they are discovered and referenced in the new AGENTS.md
- Does not add itself to any workflow chain (no contract)
- Does not need to be run again unless the project fundamentally changes
