Writing and research

Writing: you are the editor-in-chief

The productive stance is that of an editor with a fast, eager staff writer. You set the angle and the facts (the brief), the model produces the draft, you make it yours — judgment, emphasis, and the two sentences only you could write. Three habits separate pros from paste-and-pray:

• Feed it your voice. Paste two things you actually wrote and liked: 'match this register.' Otherwise you get fluent Generic Professional, and readers can tell.
• Keep your opinions yours. Have it argue for and against, then you take the position. A model-chosen position is an average, not a judgment.
• Kill the tells. Delete filler like 'In today's fast-paced world', 'It's important to note', 'delve'. If a sentence could open any document ever written, cut it.

Research: interrogate, don't transcribe

AI research goes wrong when it's treated like a search engine that returns finished truth — that's how you inherit hallucinations (the industry term for confidently made-up content). It goes right when the model is treated like a well-read colleague you interrogate: fast orientation, then verification of anything that matters.

1. 1Orient: "Explain the landscape of [topic] — the main approaches, the trade-offs, and the terms of art I should know." Perfect first move into unfamiliar territory.
2. 2Drill: ask follow-ups exactly where your understanding is thin. This beats reading ten tabs because it answers your confusion, not a generic reader's.
3. 3Extract claims: "List the five factual claims in your answer that my conclusions would depend on."
4. 4Verify those five against primary sources — the vendor's docs, the regulation's text, the paper itself. If your tool cites sources, open them; models sometimes cite pages that don't say what's claimed (or don't exist).