Before you start
Set the job up right
A few minutes of setup saves rounds of fixing later.
- Define what “done” looks like. Say who it’s for, what it must include and how you’ll check it. Otherwise it stops when the work merely looks done.
Try: “It’s done when it fits on one page and lists every step in order.” - Let it interview you. You know more than you type. Have it ask for the missing details before it writes a word.
Try: “Before you start, ask me up to five questions. Skip anything you can work out yourself.” - Show it an example of good. A sample of the tone, format or result you want often beats a paragraph describing it. Several samples work even better.
Try: “Match the style of this email:” then paste the email below it. - Get a plan before the work. For big or unclear jobs, read the plan first; skip it for small, clear ones. A wrong turn costs the least to fix on paper.
Try: “Give me a short plan first. Don’t start until I say go.”
While it works
Keep it focused and moving
- Keep the context window tight. The context window is everything the AI can see in this chat. Use one chat per job; old dead ends can make it lose track.
Try: “Sum up our goal, what we decided and what’s still open. Keep exact names and numbers.” Check it, then paste it into a new chat. - Restart instead of arguing. If two corrections haven’t fixed it, the chat is cluttered with failed tries. Start fresh with a better first message.
Try, in a new chat: “Write a 200-word bio for my work website. Use first person, and don’t mention my hobbies.” - Tell it to keep going until it’s solved. Many tools stop once the work looks done. Ask it to keep trying until a real test passes. In a plain chat app, you run the test.
Try: “Keep working until the total matches my bank statement: $1,240. If one idea fails, try the next before you check in.”
When it says done
Check the work before you trust it
- Make it prove the work. AI tools often say “done” without checking. Ask for the output, the steps or the source. Links can be made up, so open them.
Try: “How do you know this is right? Show the output or list your sources so I can check them.” - Run a critic, then a harsh one. First get a fair review. Then paste the work and goal into a new chat to find flaws. Some won’t be real.
Try: “Here is my goal and my plan. You think the plan will fail. List the biggest reasons, worst first, and say if none are serious.”
After the job
Keep what you learned
- Write down the rules you repeat. If you give the same correction in more than one chat, save it as a rule in your app’s settings. Keep the list short.
Try: “List the corrections I made today as short rules. I’ll pick which to keep.”
Where rules live.
Look in your app’s settings for custom instructions or personal preferences; projects have their own. Claude Code, a coding tool, reads a CLAUDE.md file.
Three more
Three bonus habits
Ask for a pick, not a menu. “Pick one and tell me why” gets you a decision with reasons you can check.
Give it the real thing. Attach or paste the actual file, page or data, not your summary. First remove passwords, account numbers and other private details.
Own the result. It can make up facts, quotes and sources and sound sure. Check what matters before you send it or act on it.
Showing the work
How this answer was made
Written from daily use of AI tools for building and writing, and drafted with Claude. Every claim was checked against Anthropic’s published guidance, listed below.
Then four AI critics reviewed it, one told to attack it. They caught real problems, such as advice that would backfire in a plain chat app. Those are fixed.