loop-engineer

Loop Engineer

You are not writing a prompt. You are designing a job — a feedback system that discovers, plans, executes, verifies, and iterates until the work is done, without a human manually driving each step.

Preamble — RAG pre-flight

Before starting any design work, query prior loop designs for this exact task domain. See standards/skill-patterns.md §rag-first and §mount-guard.

mount | grep -q "${DEV_ROOT}" || { echo "BLOCKED: External HD unmounted — RAG unreachable"; exit 1; }
python3 ~/.claude/rag-index/query.py "loop design <task-domain> precedent" --top 5

• If result shows a loop design for the same domain within the last 24h → surface it, ask user to confirm whether to reuse or start fresh.
• If no recent match → proceed directly to Phase 0.

Done when: RAG queried; prior loop cited or confirmed absent.

---

Output schema

Every run of this skill produces exactly these files:

goal-card.md           ← Phase 0: confirmed goal + done-when criterion
architecture.md        ← Phase 1+2: architecture + mode decision with rationale
cycle.md               ← Phase 3: 5-stage cycle mapped to THIS task concretely
building-blocks.md     ← Phase 4: 6 building blocks spec table
implementation.md      ← Phase 5: the actual runnable artifact
reconciliation-template.md ← Phase 6: the loop run output template

Save each file as you complete the phase — do not batch them at the end.

---

Six-phase workflow

Work through the phases below in order. Each phase produces a named output file. Do not skip phases or merge them.

Phase 0 — Goal Characterization Interview the user (5 questions), synthesize Goal Card, wait for user approval before Phase 1.

Phase 1+2 — Architecture & Mode Decision Decide loop size (single-agent vs. fleet) and mode (closed vs. open). Run critic gate. Confirm decisions with user before Phase 3.