Cheggin/loop-prompt

Loop Prompt Generator

Generate a loop prompt for a continuous agent. Takes an agent name, task description, and loop type. Outputs a complete prompt with iteration tracking, stop conditions, and context reset protocol.

Usage

Invoke when spawning any persistent agent. Choose the loop type that matches the work pattern.

Loop Types

1. Improvement Loop

research → change → eval → keep/discard

Best for: harness-researcher, self-improvement, skill optimization.

[IMPROVEMENT LOOP — ITERATION {{N}}/∞]

You are running an autonomous improvement loop. Each iteration:

1. READ the current state. What's the weakest area?
2. RESEARCH one improvement (web search, reference repos, knowledge wiki).
   - One variable at a time. Never change two things.
3. IMPLEMENT the change.
4. EVAL against the existing test suite / eval framework.
5. DECIDE:
   - Improved → KEEP. Commit. Log to ledger. Advance.
   - Equal perf + simpler → KEEP. Simplicity is a real win.
   - Worse or more complex → DISCARD. Git reset. Log why.
6. LOG the result to .harness/knowledge/{category}/log.md.
7. NEXT iteration. Do not pause. Do not ask.

STOP CONDITIONS:
- Plateau: <3% improvement for 4 consecutive iterations → escalate to Slack
- Budget: cost exceeds ${BUDGET} or ${MAX_TURNS} turns → stop gracefully
- Repetition: same change attempted 3 times → move to different area

CONTEXT RESET: When context fills >50%, commit all work, write handoff to GitHub Issue, start fresh session reading the Issue.

NEVER STOP. The human may be asleep. Run until interrupted.

2. Monitoring Loop

check → alert → fix → verify

Best for: ops agent, uptime monitor, error response, security scanning.

[MONITOR LOOP — TICK {{N}}]

You are running a continuous monitoring loop. Each tick:

1. CHECK all monitored systems:
   - Sentry MCP: new errors since last tick?
   - Uptime: all endpoints responding?
   - Performance: Core Web Vitals within budget?
   - Dependencies: new vulnerabilities?
2. If CLEAN: log "tick {{N}} clean" and sleep until next tick.
3. If ISSUE DETECTED:
   a. CLASSIFY: FATAL (stop everything) / TRANSIENT (retry) / UNKNOWN (investigate)
   b. ALERT: post to Slack with severity + details
   c. FIX: if within scope, diagnose → fix → deploy → verify
   d. VERIFY: confirm fix resolved the issue
   e. POST-MORTEM: log what happened, root cause, fix applied

STOP CONDITIONS:
- FATAL error requiring human input → alert Slack, pause
- Budget exceeded → stop gracefully
- Manual cancellation

NEVER STOP. Monitoring is 24/7.

3. Growth Loop

measure → hypothesize → experiment → measure

Best for: growth agent, landing page optimizer, A/B testing.

[GROWTH LOOP — EXPERIMENT {{N}}]

You are running a continuous growth optimization loop. Each experiment:

1. MEASURE current metrics (PostHog: traffic, conversion, retention).
2. HYPOTHESIZE one change that should improve the target metric.
   - State the hypothesis explicitly: "Changing X to Y should improve Z by ~W%"
   - One variable. Never test two changes at once.
3. IMPLEMENT the change (feature flag or A/B test if possible).
4. WAIT for sufficient data (minimum 100 events or 24 hours).
5. MEASURE again. Compare against baseline.
6. DECIDE:
   - Statistically significant improvement → KEEP. Ship to 100%.
   - No significant difference → DISCARD. Revert.
   - Worse → DISCARD. Revert immediately.
7. REPORT in investor update with actual numbers.

STOP CONDITIONS:
- Target metric achieved → report success
- 5 consecutive experiments with no improvement → escalate, try different strategy
- Budget exceeded

NEVER STOP. Growth compounds.

4. Maintenance Loop

scan → prioritize → fix → verify

Best for: slop-cleaner, dependency updater, tech debt reducer.

[MAINTENANCE LOOP — PASS {{N}}]

You are running a continuous maintenance loop. Each pass:

1. SCAN the codebase for the target smell:
   - Dead code, unused imports, unreachable branches
   - Duplication (copy-paste, redundant helpers)
   - Unnecessary abstractions (wrappers used once)
   - Outdated dependencies, security vulnerabilities
   - Weak or missing tests
2. PRIORITIZE: highest-impact, lowest-risk fixes first.
3. LOCK behavior: run existing tests, note what must not change.
4. FIX one smell per pass. Keep diffs small and reversible.
5. VERIFY: all tests still pass, no behavior change.
6. COMMIT with clear message describing what was removed/simplified.
7. MEASURE: lines deleted > lines added? Complexity reduced?

STOP CONDITIONS:
- No more smells found → report clean codebase
- Plateau: 3 passes with nothing found → stop
- Budget exceeded
- Test failures that can't be resolved → escalate

Prefer DELETION over REFACTORING. Simpler is always better.
NEVER STOP. Entropy is constant.

Generating a Custom Loop Prompt

When none of the templates fit exactly, compose a custom loop from these building blocks:

1. Iteration header: [LOOP_NAME — ITERATION {{N}}/{{MAX}}]
2. State check: read current state before doing anything
3. One action per iteration: never change two things
4. Eval against fixed criteria: the scorer is frozen
5. Binary decision: keep or discard, no "maybe"
6. Log everything: append to ledger/log, never modify history
7. Stop conditions: plateau + budget + repetition + manual cancel
8. Context reset protocol: commit → write handoff → start fresh
9. NEVER STOP directive: the agent runs until interrupted

Anti-Patterns

• Pausing to ask "should I continue?" — NEVER. The human set this loop running intentionally.
• Changing multiple variables per iteration — makes it impossible to know what worked.
• Skipping the eval step — "it looks better" is not evidence.
• Modifying the eval criteria mid-loop — the scorer is frozen (from autoresearch).
• Letting context fill without resetting — quality degrades past 50% context usage.