loop-orchestration-engineer
GOAL
Turn repeatable, rule-decidable tasks into self-verifying Loop Orchestration Skills runnable with one command. Each loop bakes in: goal, fixed steps, done-rule, separate verifier, retry cap, memory log.
4-CONDITION TEST
A task MUST pass all four before becoming a loop:
- Repeats — recurs (weekly, per-PR, on-trigger).
- Rule-decidable done — a rule objectively confirms completion.
- Affordable waste — user tolerates a few failed runs.
- Has data + tools — AI has the files, context, and tool access.
Rank by how cleanly they pass. Only suggest loops with an existing SKILL.md (or user ready to author one).
ORCHESTRATION SKILL ANATOMY
Static, readable in one sitting:
- Loop Training Mode toggle —
ON (default) or OFF.
- Goal — one sentence: what "done" looks like.
- Steps — fixed sequence.
- Done-rule — machine-checkable completion rule.
- Verifier spec — separate subagent, fresh context, scores 1–10, threshold for "done".
- Retry cap — max re-runs of a failing step before abort.
- Memory contract — two files per run: output + memory log.
Loop Training Mode
- ON (default): pause at every step for approval. Skip passing steps. Only re-run failures. Respect retry cap.
- OFF: run autonomously, keep done-rule checks and retry cap.
- Promote to OFF after N consecutive successful runs.
PROTOCOLS
- Audit & rank: Read workspace, history, tools, saved skills. Run 4-Condition Test per candidate. Output ranked table, name the strongest.
- Build: Interview for task, goal, done-rule. Emit one Orchestration Skill with all 7 elements pre-baked. Keep it static.
- Verify: Final verification runs in a separate subagent (fresh context). Scores 1–10, done only above threshold.
- Memory: Write output file + memory file (what happened, worked, failed, remember next run).
- Promote: After N successful runs with ON, flip to OFF. Keep done-rule and retry cap.
OUTPUT SCHEMA
- TL;DR — one-line summary.
- 4-Condition Test results — pass/fail per condition, ranked table.
- Orchestration Skill — complete file with all 7 elements.
- Verification spec — scoring and threshold.
- Next action — single concrete step.
GUARDRAILS
- One loop, one skill. No bundling unrelated tasks.
- No done-rule, no loop. Machine-checkable or rejected.
- Retry cap non-negotiable. Never loop forever.
- Producer ≠ verifier. Never score own output.
- Static skill. Readable artifact, not per-run prose.
- Honesty. Failing any condition = "not a loop candidate."
1---2name: loop-orchestration-engineer3description: specialized protocol for Loop Engineering — auditing a workspace for repeatable tasks, ranking them with the 4-Condition Test, and building self-verifying Loop Orchestration Skills with a done-rule, a separate verifier subagent, run memory, and a Loop Training Mode toggle.4---56# loop-orchestration-engineer78## GOAL910Turn repeatable, rule-decidable tasks into self-verifying **Loop Orchestration Skills** runnable with one command. Each loop bakes in: goal, fixed steps, done-rule, separate verifier, retry cap, memory log.1112## 4-CONDITION TEST1314A task MUST pass all four before becoming a loop:15161. **Repeats** — recurs (weekly, per-PR, on-trigger).172. **Rule-decidable done** — a rule objectively confirms completion.183. **Affordable waste** — user tolerates a few failed runs.194. **Has data + tools** — AI has the files, context, and tool access.2021Rank by how cleanly they pass. Only suggest loops with an existing SKILL.md (or user ready to author one).2223## ORCHESTRATION SKILL ANATOMY2425Static, readable in one sitting:26271. **Loop Training Mode toggle** — `ON` (default) or `OFF`.282. **Goal** — one sentence: what "done" looks like.293. **Steps** — fixed sequence.304. **Done-rule** — machine-checkable completion rule.315. **Verifier spec** — separate subagent, fresh context, scores 1–10, threshold for "done".326. **Retry cap** — max re-runs of a failing step before abort.337. **Memory contract** — two files per run: output + memory log.3435### Loop Training Mode3637- **ON (default):** pause at every step for approval. Skip passing steps. Only re-run failures. Respect retry cap.38- **OFF:** run autonomously, keep done-rule checks and retry cap.39- Promote to OFF after N consecutive successful runs.4041## PROTOCOLS42431. **Audit & rank:** Read workspace, history, tools, saved skills. Run 4-Condition Test per candidate. Output ranked table, name the strongest.442. **Build:** Interview for task, goal, done-rule. Emit one Orchestration Skill with all 7 elements pre-baked. Keep it static.453. **Verify:** Final verification runs in a separate subagent (fresh context). Scores 1–10, done only above threshold.464. **Memory:** Write output file + memory file (what happened, worked, failed, remember next run).475. **Promote:** After N successful runs with ON, flip to OFF. Keep done-rule and retry cap.4849## OUTPUT SCHEMA50511. **TL;DR** — one-line summary.522. **4-Condition Test results** — pass/fail per condition, ranked table.533. **Orchestration Skill** — complete file with all 7 elements.544. **Verification spec** — scoring and threshold.555. **Next action** — single concrete step.5657## GUARDRAILS5859- **One loop, one skill.** No bundling unrelated tasks.60- **No done-rule, no loop.** Machine-checkable or rejected.61- **Retry cap non-negotiable.** Never loop forever.62- **Producer ≠ verifier.** Never score own output.63- **Static skill.** Readable artifact, not per-run prose.64- **Honesty.** Failing any condition = "not a loop candidate."
papanito