Good technical writing starts with the detail you almost removed because it seemed too specific. That detail is usually the part nobody else can make up for you.
Collect the material before you write the hook
Start with a small record: what you tried, what happened, what you expected, and what surprised you. If you are explaining a paper, write down the experimental setup and one limitation. If you are describing a build, capture the command, screenshot, or test result that grounds the account.
Do not ask a model for the post before you have this material. It will happily produce an elegant narrative that accidentally turns guesses into events. Your notes create a boundary between what you know, what you infer, and what you are still asking.
Give AI an editing job, not your identity
A useful instruction is small: 'Find the sentence where I stop explaining why this benchmark result matters.' Or 'List the factual claims here that need a primary source.' That gives the model something it can do without pretending to be the author of your experience.
Ask for objections and uncertainty, then choose which ones matter. Do not accept false balance automatically. If the evidence supports a sharp conclusion, keep it. If it is ambiguous, say exactly where. Avoid letting the model flatten the voice into smooth generic phrasing just because the grammar looks clean.
1. Write raw notes from a real event, artifact, result, or question.
2. Mark each claim: observed / sourced / inferred / unknown.
3. Use AI to inspect structure, gaps, or objections.
4. Check citations and personal details yourself.
5. Read aloud and remove any sentence you would never say.Choose one thing worth saying
A post cannot carry every thought about a release. Choose the specific consequence for the reader. For a new speech model, maybe the useful question is whether it can transcribe a long English recording locally on available hardware. For an AI benchmark, maybe the issue is how a leaderboard changed the tool configuration between model versions.
Your opening can be surprising, but it should promise a payoff that the body actually delivers. Draft three versions, keep the strongest defensible one, then make sure the first paragraph identifies the subject quickly. A catchy opening that hides the topic is friction, not craft.
Edit for the human details
Cut unsupported claims, stock transitions, and filler that could be pasted into a post about any product. Keep a small concrete moment if it clarifies the mechanism. It could be an exact error message or the awkward step a benchmark setup left out. Never invent a meeting, customer, failure, quote, or personal result to make the story warmer.
End where the implication lands. A simple consequence or practical next step often carries more force than a motivational moral or a question added to harvest comments. The reader does not need to agree to receive value from the post.
You can use AI and still write something worth reading. The test is whether there is a human choice in the result about what is important and why.
FOLLOW THE EVIDENCE
Sources & further reading
Process note: AI tools assisted with research and drafting. Sources are linked so you can inspect the evidence. Unless an essay says otherwise, it does not claim hands-on testing or independent reproduction.
