Every Sunday I share one ten-minute habit that keeps my AI agents honest for the week ahead. This week’s is about the file every agent reads first: its instructions.
Each of my brand agents wakes up on a schedule and reads a rulebook before it does anything else. It says who the agent is, which tools are connected, which keys exist and what runs when. I wrote most of it in June. This morning I checked it against the actual container, one claim at a time. Seven claims had stopped being true, four blanks had never been filled in, and the agent hadn’t stopped to complain about any of them.
The receipt: 7 drifted claims and 4 blanks in my AI agent instructions
- Two tools listed as “connected” aren’t. The instructions say my task manager and my image generator are wired in as direct tool connections. Neither is. The agents found out on their own and fell back to plain API calls. Their notes files mention “not connected” 426 times. (I covered why that fallback works fine in MCP vs API.)
- One schedule rule describes a swap that doesn’t happen. The rulebook says a Monday research job replaces the daily social job. The schedule actually runs both, at different times.
- Two variables listed as available are empty. One is the music key for a video pipeline I deleted in May. The other is a brand name that a weekly analytics job is supposed to use.
- One login is described as covering three Google services. The refresh token behind it is rejected today. The job that needs it found another route, so nothing failed loudly.
- This very skill reads two project IDs from variables that were never set. Every morning it falls back to IDs typed further down the rulebook.
- And the 4 blanks: placeholders that still say “TODO”, including the booking link and the list of communities my social agent should watch. The communities placeholder alone shows up more than 200 times in the agents’ notes.
Here’s the part that matters. None of this broke a run. My agents are good at working around wrong instructions. But every workaround costs time and tokens on every run, and it hides the drift from me. A rulebook that’s 90% true teaches the agent to treat the other 10% as optional. That’s how a real rule gets skipped one day.
The 10-minute instruction drift check
- Pull out the factual claims (3 minutes). Skip the voice and tone sections. Highlight only the sentences that can be checked: “X is connected”, “Y runs at 9:00”, “Z is available”, “the link is…”. Mine is about 500 lines long. Most of it is voice and process, so the checkable claims fit on one page.
- Give each claim a one-line test (5 minutes). Is the key set (check the name, never print the value)? Does the scheduled time match the real schedule file? Does the login still return a token? Does the link load? Every claim gets a yes or a no. “Probably” is a no.
- Fix it, delete it or date it (2 minutes). If you can fix it today, fix it. If it’s dead, delete it, because a dead instruction is worse than a missing one. If it needs a decision from someone, add the date you found it next to the claim so the agent stops treating it as current.
One extra tip: search your agents’ notes for the workaround they keep writing down. “Not connected”, “unset”, “falling back to”. If the same note shows up week after week, the agent has been reporting drift the whole time. Nobody turned it into a fix.
⚡ 取得人工智慧優勢
每週提供真正省時省錢的AI小技巧。沒有廢話，沒有誇大其詞——只有切實有效的方法。.
This pairs well with two earlier Sunday checks. 10分鐘記憶力測試 looks at what the agent learned. This one looks at what you told it. And if your instructions have grown huge, 限制代理程式在啟動時讀取的內容 makes the drift check faster too.
My eleven items are on the operator list for this week. I’m not editing the rulebook in place at 7am; that call belongs to a human. Want help auditing what your agents believe about your business? 預約策略會議 and we’ll read your instructions line by line together.
下週日見。.
喬恩
📥 免費：《人工智慧劇本》
我用來經營一人代理公司的所有工具和工作流程。 25 年的行銷經驗濃縮成一份實用指南。免費贈送。.
人工智慧與數位行銷策略師
Jon Jones擁有超過25年的數位行銷、電子商務、銷售和國際商務經驗，是一位資深企業家，致力於透過創新的人工智慧解決方案賦能小型企業。他的職業生涯始於華爾街的投資銀行，曾指導企業完成資本募集和應對複雜的市場環境。如今，他的專業領域已拓展至網站開發、品牌設計、產品開發以及課程培訓。.
Jon精通普通話，並在金融和科技領域擁有深厚的功底。他創立了JonJones.ai，旨在為小型企業提供易於使用且具有變革意義的工具，幫助他們簡化營運、拓展客戶並實現永續成長。他對人工智慧、自動化和數位化策略的精通，已幫助許多企業彌合了傳統模式與現代技術之間的鴻溝，確保它們在當今的數位化環境中蓬勃發展。.
