# Feedback Loop Template **Level:** 4 **Best for:** systems that should improve over time, not just execute **Use with:** OpenClaw or any agent system that can log outcomes, re-read them on a schedule, and update its own strategy --- ## What this is Level 3 agents follow your strategy. Level 4 agents notice when the strategy is wrong and change it. That sounds like a lot. The actual mechanism is simpler than most people expect, and smaller too. Most self-improving agents are not rewriting their own code. They are doing something more like this: after every run, the agent logs what happened. On a schedule, it re-reads those logs, looks for patterns, and updates its own playbook. That's the loop. The template below gives you the formula. Six components. Without all six, it drifts. --- ## How to use this 1. **Decide what you want the system to get better at** — be specific. "Get better at growth" is not a metric. "Increase reply rate on follow-up messages" is. 2. **Fill in the log format first** — this is the most skipped step and the one that breaks the loop. If the agent doesn't log the right signals, the review has nothing to work with. 3. **Set the review cadence** — weekly is usually right for most systems. Daily is too noisy. Monthly is too slow to catch problems. 4. **Define what can change automatically and what can't** — this is the guardrail that keeps Level 4 from drifting. --- ## The formula ### 1. What gets logged (per run) This is what the agent writes to memory or a structured log after every execution. It's what the review will read. ``` RUN DATE: [date] TASK: [what the agent was doing] WHAT HAPPENED: [short factual summary — not evaluation, just what occurred] OUTCOME SIGNAL: [the metric you care about — reply rate, conversion, engagement, accuracy] WHAT WORKED: [one or two specific things that produced good signal] WHAT FAILED: [one or two specific things that produced weak or bad signal] NOTES: [anything unusual — anomaly, external factor, one-off that shouldn't influence strategy] ``` If you skip this and just log a success/fail flag, the review agent has no raw material. It can't learn from "success" alone. ### 2. The review (what the agent re-reads and asks) On a schedule, the agent pulls the last N runs and looks for patterns: ``` REVIEW WINDOW: last [7 / 14 / 30] runs QUESTIONS TO ASK: - Which conditions consistently produced strong signal? - Which conditions consistently produced weak signal? - Is there a pattern that didn't exist in the previous review? - Did anything that worked before stop working? - What should change based on this? ``` The output of the review is not a report. It is a proposed change to the strategy — specific and small. ### 3. What can be updated (the update buckets) ``` PROMPT / STRATEGY: - [what the agent is trying to do] - [tone, approach, angle] - [priorities and what gets attention first] TIMING / ROUTING: - [when the agent runs] - [which platform or channel gets focus] - [how traffic or effort is distributed] ATTENTION ALLOCATION: - [which inputs are prioritized] - [which segments, sources, or categories get more weight] ``` Updates go into one of these three buckets. If a proposed change doesn't fit, it probably needs human review, not automatic application. ### 4. What cannot be auto-updated (the guardrails) ``` LOCKED: - [core voice, tone rules, brand constraints — the things that define the work] - [any action that is irreversible: send, post, delete, charge, publish] - [anything that affects account security or access] - [any rule that exists for a reason external to performance — legal, ethical, relationship-based] ``` These never change automatically. If the review suggests touching one of these, it flags it for human review instead. ### 5. The rollback trigger ``` ROLLBACK CONDITION: If [specific metric] drops [X%] over [N review cycles], revert to the previous strategy version. Rollback is automatic. Re-activation requires human review. ``` ### 6. The review cron skill prompt This is what you give the agent that runs the weekly review. It reads the log and writes the updates back into the skill file. ``` You are the weekly review agent for [SYSTEM NAME]. Your job: read [LOG FILE] for the last [N] days and update the Self-Evolution section of [SKILL FILE]. Read the log entries and find: - Which [format / approach / channel] consistently produced [your quality signal]? - Which produced weak or no signal? - Did anything change compared to the previous review cycle? - What is one thing worth trying differently next cycle? Then rewrite these sections inside [SKILL FILE]: ### What's Working [specific patterns with evidence — name the format, the channel, the number] ### What's Not Working [specific patterns with evidence — same standard] ### Next Experiments [1–3 concrete things to try next cycle, not vague directions] ### [Your metric] Actuals (Rolling [N] Days) [real numbers vs. targets for each tracked category] Rules: - Never touch [list your locked sections] - Be specific. Vague updates ("engagement was mixed") are useless — name what drove it - If the data window is too small to draw a conclusion, say so and don't guess - Write for the agent that reads this on its next run, not for a human report ``` --- ## Example: Social Strategy Feedback Loop A social posting bot runs daily on X and Bluesky. A weekly review cron reads the posting log and rewrites four sections inside the skill file. The next run reads the updated file. The behavior changes because the spec changed. **POSTING_LOG.md — what the log actually looks like after one week:** ``` DATE: 2026-03-10 PLATFORM: Bluesky FORMAT: sequence POST_TEXT: "Claude Code from zero: Install -> first project -> 12 projects -> prompting..." ENGAGEMENT: likes=14, replies=6, follows_gained=3 QUALITY_SIGNAL: high NOTES: thread follow-up got 2 replies asking for the article link DATE: 2026-03-10 PLATFORM: X FORMAT: sequence POST_TEXT: "Claude Code from zero: Install -> first project -> 12 projects -> prompting..." ENGAGEMENT: likes=3, replies=0, follows_gained=0 QUALITY_SIGNAL: low NOTES: credit limit hit mid-day, reach may be suppressed DATE: 2026-03-11 PLATFORM: Bluesky FORMAT: tip POST_TEXT: "75 things to smoke test. Saved me from shipping 3 bugs. [link]" ENGAGEMENT: likes=9, replies=2, follows_gained=1 QUALITY_SIGNAL: high NOTES: none DATE: 2026-03-11 PLATFORM: X FORMAT: tip POST_TEXT: "75 things to smoke test. Saved me from shipping 3 bugs. [link]" ENGAGEMENT: likes=1, replies=0, follows_gained=0 QUALITY_SIGNAL: low NOTES: none DATE: 2026-03-12 PLATFORM: Bluesky FORMAT: milestone POST_TEXT: "Just hit 5,500 subscribers. Started with 0 and a 30-project challenge." ENGAGEMENT: likes=22, replies=8, follows_gained=5 QUALITY_SIGNAL: high NOTES: 3 replies from builders sharing their own milestones — engagement was genuine ``` **SKILL.md — the Self-Evolution section after the weekly cron runs:** ``` ## Self-Evolution (Last updated: 2026-03-17 by posting-weekly-review cron) ### What's Working - Bluesky: sequence and milestone formats consistently produce high-quality signal (replies from real builders, meaningful follows) - Thread-style posts on Bluesky outperform single posts 3:1 on reply quality - Milestone posts drive the highest engagement on both platforms — use sparingly - Active engagement (replies, follows) correlates with follower quality gains ### What's Not Working - X: all formats producing low-quality or zero engagement this cycle - X credit limits suppressing reach — ROI is poor relative to Bluesky effort - Promotional posts (direct product pushes without story) getting no traction anywhere ### Next Experiments - Shift primary posting effort to Bluesky: target 2 posts/day, reduce X to 1 - Test thread format more: break long tips into 2-3 connected posts - Try opening with a stat or number on Bluesky — stat-hook format hasn't been tested there yet ### Content Mix Actuals (Rolling 7 Days) - Resource sharing: 45% (target 40%) — slightly over, pull back - Building in public: 30% (target 20%) — milestone posts driving this up, OK for now - Original tips: 15% (target 20%) — slightly under - Engagement: 10% (target 10%) — on target ``` **The review cron prompt (filled):** ``` You are the weekly review agent for the Build to Launch social posting bot. Your job: read POSTING_LOG.md for the last 7 days and update the Self-Evolution section of social-posting/SKILL.md. Read the log entries and find: - Which post formats (sequence, tip, milestone, thread, question) consistently produced QUALITY_SIGNAL: high? - Which produced QUALITY_SIGNAL: low or no engagement? - Did either platform shift compared to last week's pattern? - What is one format or approach worth testing next cycle? Then rewrite these four sections inside SKILL.md under ## Self-Evolution: ### What's Working [formats and platforms producing high-quality signal — name the format, platform, number] ### What's Not Working [formats and platforms producing weak or no signal — same standard] ### Next Experiments [1–3 specific things to try next cycle — format, platform, timing, or approach] ### Content Mix Actuals (Rolling 7 Days) [real percentages per content type vs. targets in the Content Mix table] Rules: - Never touch: Voice section, Engagement Strategy approval rules, Formatting Rules, max posts per day limit, account credentials - Be specific. "Bluesky sequence posts averaged 4 replies vs. 0 on X" beats "Bluesky performed better" - If fewer than 5 log entries exist, note the thin data and make no changes to Next Experiments - Write for the bot that reads this file tomorrow morning, not for Jenny ``` **What stayed locked — never touched by the cron:** ``` LOCKED: - Voice section (tone, personality, brand rules) - Any action requiring approval: replies, quote tweets - Max posts per day hard limit - Account credentials and security settings ``` **Rollback condition:** ``` ROLLBACK CONDITION: If Bluesky follows_gained drops more than 30% over 2 consecutive review cycles, cron stops writing to Next Experiments and flags for human review instead. Autonomous updates resume only after Jenny confirms. ``` --- ## What changed between Level 3 and Level 4 here At Level 3, the agent had a strategy: post on both platforms, engage, grow. At Level 4, the agent noticed its own strategy was wrong — not because it was told, but because it read its own logs and found the pattern. The strategy update came from the loop, not from a human. The loop itself is simple. What makes it Level 4 is not the complexity. It's that the system now has the raw material to improve itself — and the guardrails to do it without drifting. --- ## Why the log format is the most important part If the agent only logs pass/fail, the review is useless. "It worked" tells the system nothing about why, so there's nothing to optimize. The log needs enough signal to answer: what specifically drove that outcome? If the agent can't answer that from the log, the loop will keep running and learning nothing. Start small. Four fields per run: what happened, what the signal was, what seemed to drive it, what didn't. That is enough to start finding patterns.