/loop-engineer

name: loop-engineer
description: >
  Domain-agnostic autonomous loop for any goal — coding, research, content, data,
  automation, or any objective. Asks 3 questions, then orchestrates a self-assembling
  agent team (resource-scout, researcher, planner, agent-factory, executor,
  verifier, auditor, memory-keeper) that researches, discovers resources, builds
  specialized agents for the goal, executes, and iterates until done. Modes:
  build (from scratch), research (investigate only), patch (fix/extend existing
  code), audit (review only, no changes). Supports persistent memory, git
  integration, and generates a completion report.
disable-model-invocation: true
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit
  - Agent
  - Glob
  - Grep
  - TodoWrite

Loop Engineer

You are running a loop engineering wizard. Follow these phases in order.

**On permission prompts:** `allowed-tools` above pre-approves Bash/Read/Write/Edit/Agent/Glob/Grep/TodoWrite for the top-level orchestrator while this skill is active — you should not be interrupted for those. Spawned subagents (via the `Agent` tool) run under the same session's permission settings, not a separate grant, so if a subagent still gets interrupted, it's the session's own `.claude/settings.json` (or global settings) that needs an allow rule, not this skill. `disable-model-invocation: true` means this only runs when explicitly invoked with `/loop-engineer` — it will never trigger itself based on conversation content.

---

> **To update loop-engineer:** `claude plugin update https://github.com/vibhasdutta/loop-engineer` (plugin install) or re-run `install.sh --update` / `install.ps1 -Update` (manual install). Updates are never applied automatically mid-loop.

---

Phase 1 — Core Wizard

Ask **one at a time**. Wait for the full answer.

**Q1 — Mode:** If the skill was invoked with an argument matching `build`, `research`, `patch`, or `audit` (e.g. `/loop-engineer patch`), use it as MODE and skip this question. Otherwise ask: > "Mode? build (new from scratch) / research (investigate and report, no code changes) / patch (fix or add a feature using the existing codebase) / audit (review existing code/output only, no changes)"

Store `MODE`. Default to `build` if the answer is unclear. See "Mode gating" under Phase 5 for how each mode changes the loop.

**Q2 — Goal:** > "What do you want the loop to accomplish? (1-2 sentences)"

After storing GOAL, generate LOOP_ID:

• lowercase, replace non-alphanumeric runs with hyphens
• Take first 4 meaningful words (skip: a, an, the, to, for, of, in, on, with, and, or)
• Max 24 chars, strip trailing hyphen
• Example: "Add authentication flow to the REST API" → "add-auth-flow-api"

Auto-set: `STOP_CONDITION` = "all tasks in loop-stack/<LOOP_ID>/PLAN.md checked", `BUDGET_STRING` = "20 turns", `MAX_TURNS` = 20.

**Q3 — Git integration:** > "Should the loop auto-commit after each verified task? (yes / no)"

Store: `MODE`, `GOAL`, `LOOP_ID`, `STOP_CONDITION`, `BUDGET_STRING`, `MAX_TURNS`, `USE_GIT`.

---

Phase 2+3 — Initialize Loop

Run the init script — creates all state files, copies agent files, and writes verifier in one command:

**Bash (macOS/Linux):**

bash ~/.claude/skills/loop-engineer/scripts/init-loop.sh \
  --loop-id <LOOP_ID> \
  --goal "<GOAL>" \
  --stop "all tasks in loop-stack/<LOOP_ID>/PLAN.md checked" \
  --git <yes/no> \
  --mode <MODE> \
  --platform claude

**PowerShell (Windows):**

& "$env:USERPROFILE\.claude\skills\loop-engineer\scripts\init-loop.ps1" `
  -LoopId "<LOOP_ID>" `
  -Goal "<GOAL>" `
  -Stop "all tasks in loop-stack/<LOOP_ID>/PLAN.md checked" `
  -Git <yes/no> `
  -Mode <MODE> `
  -Platform claude

If the script is missing, install the skill first:

• Plugin: `claude plugin install https://github.com/vibhasdutta/loop-engineer`
• Manual: `git clone https://github.com/vibhasdutta/loop-engineer && bash install.sh`

The script creates `loop-stack/<LOOP_ID>/`, `.claude/agents/` with all agent .md files + knowledge-sources/, and `verifier.md` with the actual stop condition substituted.

---

Phase 4 — Startup Sequence

**FULLY AUTONOMOUS from this point. Never pause or ask the user anything.**

Run in this order. Each step waits for completion before the next.

Step 1 — PARALLEL RESEARCHERS (dynamic count)

**Determine how many researchers to spawn based on goal complexity:**

• Simple, single-domain goal → 2 researchers
• Medium complexity or multi-domain goal → 3 researchers
• Large or multi-system goal (e.g. full-stack app, migration, complex refactor) → 4 researchers

**Spawn all researchers simultaneously in a single response** (call Agent tool N times at once).

Always assign each researcher a distinct focus area. Divide the following domains across however many you spawn — these apply to ANY goal type (coding, research, content, data, automation, etc.):

• **Context & Prior Work**: what already exists relevant to this goal — source files, documents, prior research, existing assets, related work, knowledge bases, prior loop learnings
• **External Knowledge & Resources**: what's available externally — APIs, documentation, datasets, libraries, services, reference materials, skills, MCP capabilities (use WebSearch/WebFetch)
• **Requirements & Constraints**: what must be true about the output — quality standards, format requirements, accuracy, edge cases, stakeholder needs, access restrictions, performance criteria (include this for 3+ researchers)
• **Environment & Integration**: how everything fits together — tools available, configuration, system dependencies, data flow, integration points, deployment targets (include this for 4 researchers)

For each researcher spawn:

Loop directory: loop-stack/<LOOP_ID>/
Focus: {ASSIGNED_DOMAIN} — {specific files and concerns for that domain}
GLOBAL DATA FIRST: read loop-stack/.global/MEMORY.md and loop-stack/.global/TOOLS.md.
Write findings to loop-stack/<LOOP_ID>/RESEARCH.md under "##