← ClaudeAtlas

gardenerlisted

A daily self-review loop for an AI assistant. Once a day it reads the previous day's conversations and writes a three-part report - A (results worth sharing), B (failure modes, counted by class; a candidate rule is proposed only after an A/B test shows a real lift), C (success modes the user praised, grouped by underlying principle; 2 times = pending, 6 times = solid). It only records, tests and proposes; the human approves every change. 中文:给 AI 助手用的每日复盘环(园丁):每天读一遍前一天的对话,写 A 值得发 / B 失败模式 / C 成功模式三段;失败规则先做有无规则对比测试才提,成功做法同一原则攒到 2 次待固化、6 次已固化;只记录、验证、提案,改动由人点头。
ythx-101/gardener-loop · ★ 0 · AI & Automation · score 72
Install: claude install-skill ythx-101/gardener-loop
# gardener · 园丁 Every weed you pull becomes a rule (Lauren Tan, @poteto). The gardener adds the other half: every win you repeat becomes a habit, but only after it has happened often enough and was really good, not just pleasing. ## When to use - The user wants their assistant to stop repeating corrections, or to keep doing what they liked. - The user asks for a daily / weekly review of how the assistant worked, a "self-improvement loop", or a "success mode" list. - Not for: reviewing someone else's work, or changing rules without the user's approval. ## Files - `templates/daily-review-prompt.md` - the prompt the scheduled run receives (fill the paths in `{{...}}`). - `templates/failure-modes.md`, `templates/success-modes.md` - the two running lists (one line or block per class, with count, dated examples and status). - `templates/report.md` - shape of one day's report. - `scripts/slice.py` - cut one day out of a folder of JSONL transcripts (day = 04:00 to 04:00 by default, because people work past midnight). - `scripts/gardener_eval.py` - the A/B test for a candidate rule: make exam questions from real examples, answer each with and without the rule, judge with a different model, keep an append-only exam bank, weekly regression for approved rules. - `panel/` - an animated diagram of the loop (live-panel config + the `yanyu` "word rain" view). ## Fixed procedure (one run per day) 1. **Slice.** Run `python3 scripts/slice.py --src <transcripts> --out <data>/slices` for y