npx projecta-rrr

Why I Built This

I'm a solo developer. I don't write code — Claude Code does.

Other spec-driven development tools exist; BMAD, Speckit... But they all seem to make things way more complicated than they need to be (sprint ceremonies, story points, stakeholder syncs, retrospectives, Jira workflows) or lack real big picture understanding of what you're building. I'm not a 50-person software company. I don't want to play enterprise theater. I'm just a creative person trying to build great things that work.

So I built RRR. The complexity is in the system, not in your workflow. Behind the scenes: context engineering, XML prompt formatting, subagent orchestration, state management. What you see: a few commands that just work.

The system gives Claude everything it needs to do the work and verify it. I trust the workflow. It just does a good job.

That's what this is. No enterprise roleplay bullshit. Just an incredibly effective system for building cool stuff consistently using Claude Code.

— Projecta.ai

Vibecoding has a bad reputation. You describe what you want, AI generates code, and you get inconsistent garbage that falls apart at scale.

RRR fixes that. It's the context engineering layer that makes Claude Code reliable. Describe your idea, let the system extract everything it needs to know, and let Claude Code get to work.

Who This Is For

People who want to describe what they want and have it built correctly — without pretending they're running a 50-person engineering org.

Getting Started

npx projecta-rrr

That's it. Verify with /rrr:help inside your Claude Code interface.

Installed from inside Claude Code? Slash commands won't load until you restart:

1. Type exit to quit Claude Code
2. Run claude again
3. Then run /rrr:help

This resolves "Unknown skill: rrr:help" errors.

Where to Run Commands

After installing from inside Claude Code: Exit and restart claude to load new commands.

Pushpa Mode (overnight): Always run in a separate terminal window for true unattended execution. The script detects if running inside Claude Code and prompts you to confirm or exit.

Quick Start by Scenario

MVP Definition of Done at Projecta: local demo runs + tests pass.

Pushpa Mode (Overnight Autopilot)

Run phases unattended while you sleep. Best run in a normal terminal (outside Claude Code):

bash scripts/pushpa-mode.sh
# or
npm run pushpa

Where to run: A system terminal or VS Code integrated terminal works great. Running inside the Claude interactive session can trigger approval prompts. The script detects this and prompts you to confirm or exit.

Quick guidance: /rrr:overnight — checks prerequisites and gives exact instructions.

Outputs:

• Report: .planning/PUSHPA_REPORT.md
• Logs: .planning/logs/pushpa_*.log
• Artifacts: .planning/artifacts/playwright/

Optional: Add to your project's package.json:

{
  "scripts": {
    "pushpa": "bash scripts/pushpa-mode.sh"
  }
}

Staying Updated

RRR evolves fast. Check for updates periodically:

/rrr:whats-new

Update with:

npx projecta-rrr@latest

npx projecta-rrr --global   # Install to ~/.claude/
npx projecta-rrr --local    # Install to ./.claude/

git clone https://github.com/PA-Ai-Team/projecta-rrr.git
cd projecta-rrr
node bin/install.js --local

Recommended: Skip Permissions Mode

RRR is designed for frictionless automation. Run Claude Code with:

claude --dangerously-skip-permissions

[!TIP] This is how RRR is intended to be used — stopping to approve date and git commit 50 times defeats the purpose.

{
  "permissions": {
    "allow": [
      "Bash(date:*)",
      "Bash(echo:*)",
      "Bash(cat:*)",
      "Bash(ls:*)",
      "Bash(mkdir:*)",
      "Bash(wc:*)",
      "Bash(head:*)",
      "Bash(tail:*)",
      "Bash(sort:*)",
      "Bash(grep:*)",
      "Bash(tr:*)",
      "Bash(git add:*)",
      "Bash(git commit:*)",
      "Bash(git status:*)",
      "Bash(git log:*)",
      "Bash(git diff:*)",
      "Bash(git tag:*)"
    ]
  }
}

Projecta Preferred Pack

RRR includes a default stack optimized for rapid MVP development. Defaults are recommended but overrideable.

Where Defaults Live

Core Stack (always assumed)

• Framework: Next.js (App Router) + TypeScript
• Package Manager: npm
• UI: Tailwind CSS + shadcn/ui
• Testing: Vitest (unit) + Playwright (e2e)

Preferred Providers

Agent Stack (when agents needed)

• Orchestration: Mastra
• Agent Auth: Auth.dev
• Agent Mail: Agentmail
• Sandbox: E2B
• Browser Automation: Browserbase

Overrides

If you choose a non-default provider, RRR asks for a reason and records it in Deviation Notes:

## Deviation Notes

| Capability | Default | Chosen | Reason |
|------------|---------|--------|--------|
| auth | clerk | auth0 | Client requires Auth0 for SSO compliance |

Discouraged Providers

These are allowed but require explicit justification: Firebase, Supabase, Auth0, Vercel, PlanetScale.

MCP Auto-Setup

RRR includes an MCP registry that maps your selected providers to their MCP servers. After /rrr:new-project generates your MVP_FEATURES.yml:

npm run mcp:setup

This reads your feature selections and outputs the MCP configuration for Claude Code. Only MCPs you actually need get installed.

Always included: Context7 (docs), GitHub, Filesystem, Sequential Thinking.

Pushpa Mode (Autopilot)

Run phases overnight while you sleep. Pushpa Mode is an unattended runner that plans and executes phases sequentially, with token-safe budgets and persistent state.

After running npx projecta-rrr, Pushpa Mode is automatically installed:

• scripts/pushpa-mode.sh — the runner script
• npm run pushpa — npm script shortcut

bash scripts/pushpa-mode.sh
# or
npm run pushpa

What it does:

1. Preflights required API keys based on your MVP_FEATURES.yml
2. Iterates through phases in order
3. Plans any phase that doesn't have a plan yet
4. Executes phases automatically
5. Skips phases marked with HITL_REQUIRED: true (human verification needed)
6. Runs visual proof (Playwright tests) after each phase
7. Generates a morning report at .planning/PUSHPA_REPORT.md

Token-Safe Budgets:

Pushpa Mode enforces hard limits to prevent runaway token consumption. Override via environment variables:

Persistent Ledger:

State survives crashes and restarts. The ledger at .planning/pushpa/ledger.json tracks:

• Phases started/completed
• Call counts and timing
• Failure history with signatures
• Resume point for interrupted runs

Stop Conditions:

Pushpa Mode automatically stops when:

• Budget exceeded (phases, time, or calls)
• Same failure repeats (prevents infinite retry loops)
• Git conflicts detected
• Required env vars missing

Prerequisites:

• Run /rrr:new-project first (project must be initialized)
• Set all required API keys (script will check and warn)
• Recommend enabling YOLO mode in .planning/config.json

MVP_FEATURES.yml location:

• Preferred: ./MVP_FEATURES.yml (repo root)
• Legacy: ./.planning/MVP_FEATURES.yml

Where outputs live:

• Report: .planning/PUSHPA_REPORT.md
• Logs: .planning/logs/pushpa_*.log
• Ledger: .planning/pushpa/ledger.json

HITL Convention: Plans that require human verification should include one of these markers:

• HITL_REQUIRED: true (canonical)
• HUMAN_VERIFICATION_REQUIRED
• MANUAL_VERIFICATION

Pushpa Mode will skip these phases and record them in the report for manual follow-up.

Where to run Pushpa Mode:

• Recommended: Run in a normal system terminal (outside Claude Code) for true unattended overnight runs:

  bash scripts/pushpa-mode.sh

• Running inside Claude Code works but can trigger approval prompts ("Do you want to proceed?").
• The script detects if it's running inside Claude Code and prompts: Continue running Pushpa Mode inside Claude Code? (y/N) — default is No. Press Enter to exit with instructions to run externally.

Verification Ladder

RRR uses a deterministic verification ladder with mandatory two-step verification for frontend-impacting plans:

Plan Classification:

• frontend_impact: true — plans touching frontend files, UI keywords, API routes used by UI, FE integration
• frontend_impact: false — plans with no user-visible impact (scripts, migrations, pure utilities)

Verification by Frontend Impact:

Two-Step Visual Verification (frontend_impact: true):

1. Step 1: Playwright — Automated headless E2E tests
2. Step 2: Browser UAT — Natural language browser automation via agent-browse (preferred) or claude --chrome (fallback)

Browser UAT Tool Selection:

• ANTHROPIC_API_KEY set + agent-browse plugin: Uses agent-browse (natural language browser automation via Stagehand)
• ANTHROPIC_API_KEY not set + interactive: Falls back to claude --chrome
• Pushpa/CI without ANTHROPIC_API_KEY: Skipped (no automated browser available)

Installing agent-browse plugin (in Claude Code):

/plugin marketplace add browserbase/agent-browse
/plugin install agent-browse@browserbase

Both steps run even when Playwright passes — we want visual/UI confirmation.

The planner automatically classifies plans and adds verification: frontmatter:

verification:
  surface: ui_affecting
  frontend_impact: true
  required_steps:
    - unit_tests
    - playwright
    - browser_uat

Visual Proof (Mandatory Two-Step Verification)

RRR includes automated visual proof via Playwright tests and browser UAT verification.

What it captures:

• Playwright test results (pass/fail)
• Browser UAT confirmations (page renders, interactions, layout)
• Console errors and warnings
• Page errors (uncaught exceptions)
• Network failures (4xx, 5xx, failed requests)
• Screenshots, traces, and videos (on failure)

Artifacts location:

.planning/
├── VISUAL_PROOF.md                    # Append-only run log (both steps)
└── artifacts/
    ├── playwright/
    │   ├── report/                    # HTML report (index.html)
    │   │   └── index.html
    │   └── test-results/              # Per-test artifacts
    │       ├── screenshot.png         # Captured on failure
    │       ├── trace.zip              # Recorded trace
    │       └── video.webm             # Recorded video
    └── browser-uat/
        └── {run_id}/                  # Browser UAT artifacts
            ├── uat.log                # Execution log
            ├── screenshot-*.png       # Captured screenshots
            └── chrome-notes.md        # Claude chrome notes (fallback mode)

Modes (configured in .planning/config.json):

Browser UAT Gating (in .planning/config.json):

{
  "visual_proof": {
    "uat_gate": "non_blocking"  // or "blocking"
  }
}

• non_blocking (default): UAT failures logged but don't block execution
• blocking: UAT failures exit with code 1 (for strict enforcement)

Commands:

# Playwright
npm run e2e              # Run Playwright tests (headless)
npm run e2e:headed       # Run with browser visible
npm run e2e:ui           # Playwright UI mode (interactive)
npm run visual:open      # Open last HTML report

# Visual Proof (two-step for frontend_impact: true)
bash scripts/visual-proof.sh           # Playwright only
bash scripts/visual-proof.sh --chrome  # Playwright + browser UAT (REQUIRED for FE)
bash scripts/visual-proof.sh --pushpa  # Pushpa mode (headless)

# Browser UAT directly
FRONTEND_IMPACT=true bash scripts/browser-uat.sh  # Run browser UAT step

Browser UAT Tool Selection:

• agent-browse (preferred): Requires ANTHROPIC_API_KEY and browserbase/agent-browse plugin. Uses Stagehand for natural language browser automation.
• claude --chrome (fallback): Interactive mode only. Falls back when no API key set.
• Skipped: In Pushpa/CI without ANTHROPIC_API_KEY (no automated browser available).

Installing agent-browse (in Claude Code):

/plugin marketplace add browserbase/agent-browse
/plugin install agent-browse@browserbase

Interactive verification (UI_AFFECTING plans only):

npx playwright test --ui     # Playwright UI mode for exploratory testing
npx playwright test --headed # See tests execute in browser
FRONTEND_IMPACT=true bash scripts/browser-uat.sh  # Browser UAT check

Integration:

• Per-plan: Runs automatically after /rrr:execute-plan when frontend_impact: true (Playwright + Browser UAT)
• Per-phase: Runs automatically after /rrr:execute-phase (aggregates all plans)
• Pushpa Mode: Headless Playwright + agent-browse (if ANTHROPIC_API_KEY set); no interactive fallback
• Backend-only plans (frontend_impact: false): Unit tests only, visual steps skipped
• Results logged to .planning/VISUAL_PROOF.md (append-only)

VISUAL_PROOF.md format: Each run appends an entry with datetime, plan/phase ID, surface type, commands run, pass/fail status, console errors, and artifact paths.

UX Telemetry Fixture:

Bootstrap includes a telemetry fixture at e2e/fixtures/ux-telemetry.ts. Import it in your tests to capture browser events automatically:

import { test } from '../fixtures/ux-telemetry';

test('my test', async ({ page }) => {
  // Console errors, page errors, and network failures
  // are automatically captured during this test
});

How It Works

1. Initialize Project (~10 minutes)

/rrr:new-project

One command, one flow. The system:

1. Questions — Asks until it understands your idea completely (goals, constraints, tech preferences, edge cases)
2. Research — Spawns parallel agents to investigate the domain (optional but recommended)
3. Requirements — Extracts what's v1, v2, and out of scope
4. Roadmap — Creates phases mapped to requirements

You approve the roadmap. Now you're ready to build.

Creates: PROJECT.md, REQUIREMENTS.md, ROADMAP.md, STATE.md, .planning/research/

2. Plan Phase

/rrr:discuss-phase 1   # Optional: clarify UI/UX/behavior decisions first
/rrr:plan-phase 1

discuss-phase (optional) — If the phase has gray areas (UI choices, UX flows, behavior decisions), discuss them first. Creates CONTEXT.md that guides planning. Skip if you trust the system's defaults.

plan-phase — The system:

1. Researches — Investigates how to implement this specific phase
2. Plans — Creates 2-3 atomic task plans with XML structure
3. Verifies — Checks plans against requirements, loops if needed

Ready when plans pass verification.

3. Execute Phase

/rrr:execute-phase 1

The system:

1. Runs plans in waves — Parallel where possible, sequential when dependent
2. Fresh context per plan — 200k tokens purely for implementation, zero degradation
3. Verifies code — Checks against phase goals when complete

4. Repeat

/rrr:plan-phase 2
/rrr:execute-phase 2
...
/rrr:complete-milestone   # When all phases done

Loop plan → execute until milestone complete. Ship your MVP. Start next milestone.

Existing Projects (Brownfield)

Already have code? Start here instead.

1. Map the codebase

/rrr:map-codebase

Spawns parallel agents to analyze your code. Creates .planning/codebase/ with structured analysis of your stack, architecture, conventions, and concerns.

2. Initialize and build

/rrr:new-project

Same flow as greenfield, but the system knows your codebase. Questions focus on what you're adding/changing. Then plan → execute as normal.

The codebase docs load automatically during planning. Claude knows your patterns, conventions, and where to put things.

Why It Works

Context Engineering

Claude Code is incredibly powerful if you give it the context it needs. Most people don't.

RRR handles it for you:

Size limits based on where Claude's quality degrades. Stay under, get consistent excellence.

XML Prompt Formatting

Every plan is structured XML optimized for Claude:

<task type="auto">
  <name>Create login endpoint</name>
  <files>src/app/api/auth/login/route.ts</files>
  <action>
    Use jose for JWT (not jsonwebtoken - CommonJS issues).
    Validate credentials against users table.
    Return httpOnly cookie on success.
  </action>
  <verify>curl -X POST localhost:3000/api/auth/login returns 200 + Set-Cookie</verify>
  <done>Valid credentials return cookie, invalid return 401</done>
</task>

Precise instructions. No guessing. Verification built in.

Subagent Execution

As Claude fills its context window, quality degrades. You've seen it: "Due to context limits, I'll be more concise now." That "concision" is code for cutting corners.

RRR prevents this. Each plan is maximum 3 tasks. Each plan runs in a fresh subagent — 200k tokens purely for implementation, zero accumulated garbage.

No degradation. Walk away, come back to completed work.

Atomic Git Commits

Each task gets its own commit immediately after completion:

abc123f docs(08-02): complete user registration plan
def456g feat(08-02): add email confirmation flow
hij789k feat(08-02): implement password hashing
lmn012o feat(08-02): create registration endpoint

[!NOTE] Benefits: Git bisect finds exact failing task. Each task independently revertable. Clear history for Claude in future sessions. Better observability in AI-automated workflow.

Every commit is surgical, traceable, and meaningful.

Modular by Design

• Add phases to current milestone
• Insert urgent work between phases
• Complete milestones and start fresh
• Adjust plans without rebuilding everything

You're never locked in. The system adapts.

Commands

Core Workflow

Navigation

Verification

Brownfield

Phase Management

Milestones

Session

Utilities

Skills

Skills System

Skills are context packages that load into Claude during plan execution. They enforce coding conventions, tool choices, and best practices.

How Skills Work

1. Planning — Planner infers which skills are needed based on phase content
2. Execution — Skills load into executor context (fresh per plan)
3. Enforcement — Claude follows skill rules during implementation

Skills are loaded dynamically per plan. No skill accumulation across plans — keeps context clean.

Vendored Skills

Ships with RRR:

Projecta Skills (default stack):

• projecta.nextjs-typescript ⭐ — Next.js App Router + TypeScript conventions (loads automatically)
• projecta.testing — Vitest + Playwright (NEVER Jest/Cypress)
• projecta.visual-proof — Visual proof artifact capture
• projecta.shadcn-ui — shadcn/ui components (NEVER Material UI/Chakra)
• projecta.cloudflare-r2 — Cloudflare R2 storage conventions
• projecta.mcp-stack — MCP server integrations

Anthropic Skills (vendored from upstream):

• anthropic.pdf — PDF document handling
• anthropic.xlsx — Excel spreadsheet handling
• anthropic.webapp-testing — Web application testing
• And 13 more...

Using Skills in Plans

Skills are auto-inferred from plan content, but you can declare them explicitly:

---
phase: 04-auth
plan: 01
skills:
  - projecta.testing
  - projecta.visual-proof
  - anthropic.webapp-testing
---

Default skill: projecta.nextjs-typescript loads automatically for ALL plans. Add skills_mode: minimal to skip it (rare, for non-standard stacks).

Limits: Max 10 skills per plan, max 1000 total lines.

Community Skills

Install skills from GitHub or skillsmp.com:

/rrr:install-skill https://github.com/user/repo/blob/main/SKILL.md
/rrr:search-skills react patterns

Community skills install to .claude/skills/community/.

Vendoring Skills (Maintainers)

RRR uses vendoring scripts to pull skills from external repositories into the rrr/skills/ directory:

Anthropic upstream skills (scripts/vendor-skills.sh): Vendors official Anthropic skills into rrr/skills/upstream/anthropic/.

Vercel Agent Skills (scripts/vendor-vercel-skills.sh): Vendors curated React/Next.js best practices from vercel-labs/agent-skills. Small and included in npm package.

droid-tings Skills (scripts/vendor-droid-tings-skills.sh): Vendors AI/ML focused skills from ovachiever/droid-tings. Large (~40MB), not included in npm to keep package lean. Vendor on-demand when needed.

# Vendor Anthropic skills (maintainer use)
bash scripts/vendor-skills.sh

# Vendor Vercel skills (small, recommended for frontend)
bash scripts/vendor-vercel-skills.sh

# Vendor droid-tings skills (large, use allowlist)
bash scripts/vendor-droid-tings-skills.sh --list                    # List available
DROID_TINGS_ALLOWLIST="axolotl,unsloth" bash scripts/vendor-droid-tings-skills.sh  # Specific skills

npm Package Policy:

• projecta/*, upstream/*, community/vercel/* — included in npm
• community/droid-tings/* — excluded (40MB), vendor locally if needed

Troubleshooting

Commands not found after install/update?

• Restart Claude Code: type exit, then run claude again
• Verify files exist in ~/.claude/commands/rrr/ (global) or ./.claude/commands/rrr/ (local)
• Still seeing "Unknown skill"? Reinstall with npx projecta-rrr@latest, then restart claude

Commands not working as expected?

• Run /rrr:help to verify installation
• Re-run npx projecta-rrr to reinstall

Updating to the latest version?

npx projecta-rrr@latest

Using Docker or containerized environments?

If file reads fail with tilde paths (~/.claude/...), set CLAUDE_CONFIG_DIR before installing:

CLAUDE_CONFIG_DIR=/home/youruser/.claude npx projecta-rrr --global

This ensures absolute paths are used instead of ~ which may not expand correctly in containers.

Star History

License

MIT License. See LICENSE for details.