# AI Writing Policy > [!abstract] The one-line version > **Write your own thinking. Flag anything the machine wrote. Never make someone read something nobody bothered to write.** We are not anti-AI. We use it heavily. This policy exists so that when you read something from a colleague, you know exactly what you are reading — a person's thinking, or a machine's output someone thought was worth showing you. --- ## Principles **1. The writing *is* the thinking.** Deciding what to emphasise, what to cut, and how to order it is where the understanding actually happens. If you hand that step to a model, you skip the work the document existed to do. You end up with a deliverable and no insight. **2. Protect independent thought.** Ideas that come out of your own head, wrestled into shape, are the only thing here that compounds. Models converge — everyone prompted the same way gets the same answer. Our edge is the thinking that doesn't come out of a model. See [[Fluid Thinking]]. **3. You own every sentence you send.** "That bit was AI" is not a defence. If it goes out under your name, you stand behind it. If you would not defend it in a meeting, do not send it. **4. Reader time is the scarce resource.** More time should go into writing a document than reading it — one author, many readers, so the effort multiplies. A 2,000-word output from a 20-word prompt inverts that and pushes your work onto everyone else. See [[Writing Better]]. **5. Short beats polished.** AI produces fluent, confident, empty prose very cheaply. Length is not effort. Cut it. --- ## Rules > [!important] The separation rule > **Human writing and AI output never share a page.** Your thinking goes in the main doc. Anything generated goes on a **separate tab, appendix, or clearly fenced block** — labelled, so nobody spends attention on it by accident. 1. **Separate the tabs.** Main tab = what you wrote and thought. Second tab = raw AI output, research dumps, agent runs. Same for docs (appendix), Slack (thread, not channel), and email (below the fold). 2. **Label it plainly.** `[AI-generated]`, `[AI draft — not reviewed]`, or `Claude suggested this — worth exploring?`. Say which part, not just that AI was "involved". 3. **Unlabelled = yours.** If it isn't flagged, we will all assume you wrote it, thought it, and will defend it. That is the whole contract. 4. **Never send unreviewed output as a finished thing.** Sending raw generation and letting the reader sift it is transferring your work to them. 5. **Use AI freely on the way in.** Research, summarising, critique, first drafts you then rewrite, editing your own words, code, formatting. None of that needs flagging — you did the thinking. 6. **Flag it on the way out.** The moment a reader could mistake generated prose for your reasoning, it gets labelled. 7. **Think first, prompt second.** For anything that matters — a decision, a strategy, a proposal — write your own rough version before you open a model. Then use AI to sharpen it, not to skip it. 8. **Agents write as agents.** Anything an agent sends on your behalf is identified as an agent. A human name on machine output is misattribution. --- ## Why this matters Reading something generated that nobody thought about wastes real time — and it is quietly disrespectful of the reader's attention. The flip side is the point: when we know a document is a person's actual thinking, we can argue with it properly, build on it, and trust it. Related: [[Operating Principles]] · [[On Writing]] · [[Executive Operating System MOC]] --- > [!quote] Stick this bit on the wall > Write it yourself. Flag the machine. Keep it short.