Loop Maker — Skill Setup¶
Source: EricTechPro/loop-maker Stars: 0 ⭐ | License: MIT Created: June 24, 2026 Category: Agent Infrastructure / Loop Engineering / Automation Cross-Harness: Claude Code · Codex · Hermes · OpenClaw
1. What It Is¶
A portable agent skill that interviews you through a 7-question wizard and scaffolds a self-running autonomous loop — complete with separate verifier, state file, stop conditions, and human gate. Works across all major agent harnesses.
Why most loops fail (and how this catches it): 1. ❌ Agent judges its own work → ✅ Separate verifier built into scaffold 2. ❌ State leaks into the skill file → ✅ Dedicated STATE.md outside the skill 3. ❌ No stop condition → ✅ Stop condition wired in before you run
2. Prerequisites¶
- Any supported agent harness (Hermes Agent, Claude Code, Codex CLI, or OpenClaw)
- No additional packages required
3. Installation¶
git clone https://github.com/EricTechPro/loop-maker.git
cp -r loop-maker ~/.hermes/skills/loop-maker
Verify¶
ls ~/.hermes/skills/loop-maker/
# Should show: SKILL.md
4. The 7-Question Wizard¶
When invoked, loop-maker asks these questions one at a time:
| # | Question | What It Catches |
|---|---|---|
| 1 | Goal — What checkable condition means "done for now"? | Vague/unfalsifiable goals |
| 2 | Trigger — What starts each run? Schedule, event, or run-until-done? | Undefined activation |
| 3 | Discovery — How does it find the work to do each round? | Blind loops with nothing to do |
| 4 | Action — What's it allowed to do, and through which tools? | Unbounded capabilities |
| 5 | Verification — Who checks the result, and against what? | Agent self-grading |
| 6 | State — Where does "what's done / what's left" live? | Amnesia between runs |
| 7 | Human gate — Which actions are irreversible and must ask first? | Catastrophic automation |
Output¶
loop-output/
├── SKILL.md # The loop skill definition
├── STATE.md # Persistent state file (outside the skill)
├── verifier.py # Separate verification script
└── README.md # Runbook for the loop
5. Usage¶
# Start the loop wizard
hermes -s loop-maker "I need a loop that..."
# The skill interviews you through the 7 questions
# Then emits a ready-to-run loop folder
6. Comparison: Loop Maker vs Perfectloop¶
| Feature | Loop Maker | Perfectloop |
|---|---|---|
| Questions | 7 (goal through human gate) | 8 (adds verifier-verification + scheduler cleanup) |
| Cross-harness | ✅ Claude Code, Codex, Hermes, OpenClaw | Hermes-specific |
| Output | Folder: SKILL.md + STATE.md + verifier.py | YAML loop spec |
| Interview style | Conversational wizard | Formal 8-question contract |
| Rejection | Scaffolds anyway (warns) | Flatly rejects bad candidates |
| Best for | Quick loop prototyping, cross-platform | Production-critical loops, safety-first |
7. CorpusIQ Use Cases¶
| Automation | Loop Maker Application |
|---|---|
| Daily skills sweep | Trigger: cron; Verify: GitHub API returns valid; State: last_discovered_id |
| Email watchdog | Trigger: poll interval; Verify: SMTP delivery confirmed; Human gate: auto-reply content |
| Lead qualification pipeline | Trigger: new email; Verify: human reviews first 5; State: qualified_leads.json |
| Competitive monitoring | Trigger: daily; Verify: diff against previous run; State: last_snapshot.json |
8. Troubleshooting¶
| Symptom | Cause | Fix |
|---|---|---|
| Loop output missing verifier | Interview skipped question 5 | Re-run; question 5 is mandatory |
| STATE.md grows without bound | No cleanup defined | Add max_entries or ttl_days to state management |
| Loop runs forever | Stop condition too vague | Use a concrete, checkable condition (count, timestamp, boolean) |
| Cross-harness compatibility | Tool names differ | Loop Maker uses abstract tool references — bind to your harness |