Stop writing better prompts. Start writing shift notes. My agents kept failing… · @ReignSwords · TwKit

@ReignSwords

Stop writing better prompts. Start writing shift notes. My agents kept failing the same two ways. They’d try to eat the whole project in one sitting, run out of context mid-bite, and hand the next shift a half-built mess. Or they’d see some progress and declare victory at 40%. Anthropic’s engineering team wrote up their fix for exactly this. It’s not a new model. It’s a handover protocol: two prompts and a text file. Steal it 👇 📋 Prompt 1 — the Initializer (run ONCE): “Don’t build anything. Set up: http://init.sh that boots the dev server and runs one smoke test, features.json with every feature marked failing, claude-progress.txt, and one git commit.” Use JSON, not Markdown — agents corrupt Markdown. 📋 Prompt 2 — the Coder (every session): “Do exactly ONE feature. Then leave it merge-clean: commit with a real message, flip the feature’s passes flag, append to the progress file.” And steal their one hard rule verbatim: “It is unacceptable to remove or edit tests.” 📋 The session-start ritual (every run): pwd → read git log + progress file → pick the top failing feature → run http://init.sh first, fix the broken state BEFORE building anything new. Why it works: your agent starts every shift with amnesia. The text file is its memory. The JSON is its to-do list. The git log is its alibi. Fellow AI builders — follow for more like this. My new rule of thumb: if my agent can’t pick up the project cold from the progress file, the file is the bug, not the agent. Where do your agents usually lose the thread — the mid-feature amnesia, or the 40% victory lap? 👇 🔵 Fellow Premium members, let’s connect! I post about AI & tech — follow for more 👥

曝光 69 · 评论 6 · 点赞 1 · 书签 0