Iteration: refining to perfection

Beginners try to nail the perfect prompt on the first try. People who do this for a living work in loops: draft → targeted edit → edit → done. It's faster and it's more reliable.

Targeted edits beat "redo it"

"Redo it" is a lottery — the model will change the good parts along with the bad. Say what to change and what to leave alone:

• "Second paragraph is great, don't touch it. Make the first one drier and shorter"
• "Tone is right. Swap the coffee example for something from tech"

Make the model critique itself

One of the strongest moves you have — ask it to audit its own work:

• "Find 3 weak spots in your answer and fix them"
• "Score this answer on accuracy, clarity and completeness. What would you improve?"
• "What questions will the reader still have after this?"

Ask for options, not an answer

When to start over

If three or four edits in the answer keeps drifting further from what you want, the chat's context is clogged with contradictory instructions. Don't fight it: open a new chat with a better prompt that includes everything the iterations taught you — and the very first draft will land where the old chat's fourth edit did.

What it looks like live: four moves to done

You: [prompt by the formula: role, task, context, format] → a LinkedIn post about hiring
AI: [solid draft, but a stock opening paragraph]
You: The opening is a cliché — "in today's world". Start with the candidate story from my context.
AI: [rewrites — alive now, but the call to apply is gone]
You: Great, lock that in. Bring back the closing call from the first version, word for word.
AI: [assembles the final from the best of both]

Three edits, two minutes. Notice the moves: name the exact spot ("the opening paragraph"), protect what works ("lock that in"), reuse the good bits from earlier versions ("bring back the one from version one"). A chat is an editing room, not a roulette wheel.

The rule of three iterations

Thousands of users converge on the same rhythm: a draft plus two or three targeted edits gets you 90% of the available quality. After that, diminishing returns kick in hard — the fifth and seventh edits eat your time for microscopic gains. If it's still "close, but not it" after the third round, the problem isn't the edits, it's the original prompt: something's missing from the context or the format. Go back and strengthen it.

Five edit lines for everyday use

Save these — they cover most situations:

1. "Cut this by a third, keep every number and example" — for long but useful.
2. "Two notches drier/warmer in tone, leave everything else" — mood calibration.
3. "Swap the example from [X] for one from my field: [Y]" — someone else's examples become yours.
4. "Add structure: subheads and bullets, don't cut the text" — from wall of text to document.
5. "Strengthen the weakest part — and tell me why it was weak" — an edit that teaches: the model explains what it fixed.

Notice number five: asking the model to comment on its edits is a fast way to learn editing yourself. A month of those comments and you'll start spotting the weak spots before you even send the prompt.

Do this now

Grab any mediocre answer from your chat history and take it to good with three targeted edits, using the pattern from this lesson: exact spot → what changes → what stays. Time yourself. It usually comes in under three minutes, which is the best argument there is for always doing it this way.

Editing by voice

Edits work just as well in voice mode (you'll remember it from part 2 of the track): "hold on, that's too formal — say it plainly", "go back to the second version and put the numbers in". People polish whole emails and plans this way while walking, never touching a keyboard. The mechanics are identical to typing: name the spot, say what changes, protect what already works. Voice actually speeds the loop up — saying an edit is quicker than typing it, and the model follows spoken instructions just as precisely.