A community-driven registry for Claude, Cursor, Windsurf, Cline & more. Not affiliated with Anthropic.
Are you the author? Sign in to claim
A Claude Code skill that turns Claude into an infinite development loop: spec → feature list → implement → test → commit
A Claude Code skill that turns Claude into an infinite development loop: spec → feature list → implement → test → commit → repeat — across unlimited context windows.
Based on Anthropic's research: Effective harnesses for long-running agents
┌─────────────────────────────────┐
│ PROJECT FILES │
│ │
│ AGENTS.md feature_list.json │
│ init.sh claude-progress.txt│
│ dev-agent.py src/ │
└───────────────┬─────────────────┘
│
▼
┌───────────────────────────────────────────────────────────────┐
│ FIRST SESSION │
│ │
│ User Spec ──► Initializer Agent │
│ │ │
│ ├── Generate feature_list.json (20-200) │
│ ├── Create init.sh (idempotent) │
│ ├── Initialize git repo + scaffold │
│ ├── Write claude-progress.txt │
│ └── Ask user: Mode A or Mode B? │
└───────────────────────────────┬───────────────────────────────┘
│
▼
┌───────────────────────────────────────────────────────────────┐
│ PER-FEATURE WORKFLOW │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ STEP 1 │──►│ STEP 2 │──►│ STEP 3 │──►│ STEP 4 │ │
│ │ Init Env │ │ Pick │ │ Write │ │ Verify │ │
│ │ ./init.sh│ │ Feature │ │ Code │ │ │ │
│ └──────────┘ └──────────┘ └──────────┘ └────┬─────┘ │
│ │ │
│ ┌──────────────────┤ │
│ │ │ │
│ ┌────▼─────┐ ┌────────▼──────┐ │
│ │ npm run │ │ Browser Test │ │
│ │ lint │ │ (Playwright) │ │
│ └────┬─────┘ └────────┬──────┘ │
│ │ │ │
│ ┌────▼─────┐ │ │
│ │ npm run │ │ │
│ │ build │ │ │
│ └────┬─────┘ │ │
│ │ │ │
│ └────────┬─────────┘ │
│ │ │
│ ┌─────▼─────┐ │
│ │ All pass? │ │
│ └─────┬─────┘ │
│ ┌────────────┼────────────┐ │
│ │ NO │ YES │ │
│ ▼ │ ▼ │
│ ┌──────────┐ │ ┌──────────┐ │
│ │ Fix │ │ │ STEP 5 │ │
│ │ Errors │───────┘ │ Complete │ │
│ └────┬─────┘ │ + Log │ │
│ │ stuck? └────┬─────┘ │
│ ▼ │ │
│ ┌──────────┐ ┌─────▼─────┐ │
│ │ Skip + │ │ STEP 6 │ │
│ │ Revert │ │ Commit │ │
│ └────┬─────┘ └─────┬─────┘ │
│ │ │ │
│ └────────────┬───────────┘ │
│ │ │
│ ◄── next feature ──► │
└──────────────────────────────────┬────────────────────────────┘
│
┌───────────────┼───────────────┐
▼ ▼
┌───────────────────────┐ ┌───────────────────────┐
│ MODE A: Interactive │ │ MODE B: Autopilot │
│ │ │ │
│ Claude implements │ │ dev-agent.py run │
│ directly in session │ │ spawns claude -p │
│ │ │ per feature │
│ /clear + "go ahead" │ │ │
│ to reset context │ │ auto context reset │
│ │ │ fully unattended │
│ ✓ MCP/Playwright │ │ │
│ ✓ No special perms │ │ ✗ Needs --danger.. │
│ ✓ See every step │ │ ✗ MCP may not work │
│ △ Occasional /clear │ │ ✗ Can't see live │
└───────────────────────┘ └───────────────────────┘
State Persistence (across all context windows):
feature_list.json ◄── source of truth, only "passes" flips
claude-progress.txt ◄── structured session notes (What/Testing/Notes)
git history ◄── code changes + descriptive commits
npx skills add RobertWang4/infinite-dev-skill
Or manually copy to ~/.claude/skills/infinite-dev/.
npx skills update
This updates all installed skills to their latest versions, not just this one.
Just tell Claude what to build — everything else is automatic:
> Build me a task management app with React and Express
# Claude auto-triggers this skill:
# 1. Copies dev-agent.py + AGENTS.md to your project
# 2. Reads your spec → generates feature_list.json
# 3. Asks you: Mode A (interactive) or Mode B (autopilot)?
# 4. Starts building features one by one
Claude implements features in your current session. You /clear between batches.
┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ Claude │────►│ init + │────►│ Pick + │────►│ Test + │
│ reads │ │ status │ │ Code │ │ Commit │
│ AGENTS.md │ │ │ │ │ │ │
└──────────┘ └──────────┘ └──────────┘ └─────┬────┘
▲ │
│ ┌────────────────────────┐ │
│ │ Context getting long? │◄───────────┘
│ └───────────┬────────────┘
│ NO │ │ YES
│ ▼ ▼
│ ┌─────────┐ ┌──────────────────────────┐
│ │ Next │ │ "Done N features. │
│ │ feature │ │ /clear then go ahead." │
│ └────┬────┘ └────────────┬─────────────┘
│ │ │
│ │ User: /clear
│ │ User: go ahead
│ │ │
└───────────────────┴────────────────────┘
dev-agent.py run spawns a fresh claude -p process per feature. No human interaction needed.
┌──────────────────────────────────────────────────────┐
│ python dev-agent.py run │
│ │
│ while features remaining: │
│ ┌─────────────────────────────────────────────┐ │
│ │ claude -p (fresh context, fresh process) │ │
│ │ │ │
│ │ implement → lint → build → test │ │
│ │ → complete → commit → log → exit │ │
│ └─────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ context destroyed → next feature │
│ │
│ done → final report │
└──────────────────────────────────────────────────────┘
python dev-agent.py run # Use CLI default model
python dev-agent.py run --max-features 10 # Limit per run
python dev-agent.py run --timeout 3600 # 1 hour per feature
python dev-agent.py run --model <name> # Override model
[
{
"id": 1,
"category": "functional",
"priority": 1,
"description": "User can create a new chat and send a message",
"steps": [
"Navigate to main interface",
"Click 'New Chat' button",
"Type a message",
"Press Enter",
"Verify AI response appears"
],
"depends_on": [],
"passes": false
}
]
priority: 1 = highest, lower numbers go firstdepends_on: feature IDs that must pass firstpasses: false → true (passing) / "skipped" (blocked)| Command | Description |
|---|---|
dev-agent.py run | Autopilot: one claude -p per feature |
dev-agent.py status | Show progress (passing/total/skipped) |
dev-agent.py next | Show next feature to implement |
dev-agent.py complete <id> | Mark feature as passing |
dev-agent.py skip <id> "reason" | Skip a blocked feature |
dev-agent.py regression | Pick 1-2 passing features to re-verify |
dev-agent.py log --feature-id <id> --done "..." --testing "..." --notes "..." | Structured progress log |
| Project Type | Testing Method |
|---|---|
| Web apps | Playwright MCP — browser automation + screenshots |
| CLI tools | pytest / jest / go test / cargo test |
| APIs | Integration tests + endpoint verification |
| All types | Regression testing of previously passing features |
For maximum throughput, run multiple features in parallel:
Claude (Team Lead)
├── git worktree A → Feature #3 (Agent 1)
├── git worktree B → Feature #4 (Agent 2)
├── git worktree C → Feature #5 (Agent 3)
└── Merge all → Regression test → Continue
infinite-dev/
├── SKILL.md # Skill entry point (Claude reads this)
├── README.md # This file
├── scripts/
│ └── dev-agent.py # CLI: state management + autopilot loop
└── templates/
└── AGENTS.md # Template → copy to project root
MIT
1000+ skills curated from Anthropic, Vercel, Stripe, and other engineering teams
Claude Code skill for YouTube creators — channel audits, video SEO, retention scripts, thumbnails, content strategy, Sho
Design enforcement with memory — keeps your UI consistent across a project
AI image generation skill for Claude Code -- Creative Director powered by Gemini