Your first real task — iterated
Build-along · ~10 min at your own keyboard
First drafts from AI are often close but not right. The skill is judging a draft against what you needed, saying exactly what's wrong, and knowing when to stop. Before you ask, write down two or three things the result must do. Then check each draft against that list. Revise only where it misses. And if an answer loses you, don't nod along. Say 'Wait, I don't follow. Re-explain that in plain words, from the start.' Asking it to re-pitch costs one message; building on something you didn't understand costs much more.
The long version: Iterating with Claude Code ↗
Do this
0/6Checkpoints
- ▸You should now see a first draft of your task in the reply.
- ▸You should now see which of your checks it passes and which it misses.
- ▸You should now see a final version that passes every check, or a written reason it already did.
The three ways this goes wrong
- Feedback like 'make it better' gives you a slightly different but equally weak result. — Name the exact fix: which paragraph, what length, what tone. One concrete change per round.
- Claude starts editing files you did not mean it to touch. — Press Esc to stop it, then say which file is in scope. Keep the task inside one folder.
- You accept round 1 because it looks fine, without checking. — 'Looks fine' is not a check. Go through your list line by line. If it really passes, that's a good result — write down why.
Where these ideas come from: The wait-what skill (Matt Pocock's skills (MIT)). Paraphrased here.
Share what happened
Paste: the task, your checks, any feedback lines you gave, and one sentence on how the final version measures against the checks.
Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].
🔒 Don't paste confidential, personal-health, or payment info.