The second prompt matters most

Watch someone who finds these tools useless and someone who finds them valuable, and the visible difference is short: the first person reads the answer and closes the tab. The second person replies.

The first answer is diagnostic

You cannot see your own prompt from outside your head. You know what you meant, so you can’t tell what you actually said. The first answer solves this — it shows you how your request read to something with no access to your intentions.

If it came back too formal, you didn’t specify tone. Too long, you didn’t specify length. About the wrong thing entirely, your task verb was ambiguous or your context was missing. That is not a wasted round; that’s the checklist from L2 telling you which box was empty.

Say why, not just what

The instinct is to reject and re-ask: “no, try again.” That reliably produces a variation on the thing you just rejected, because nothing told it what was wrong.

Compare:

• “Shorter.” → shorter, still wrong in the same way.
• “Cut it to 150 words. The problem is the opening — it apologises before saying what we’re doing about it, and that’s what made the customer escalate last time.”

The second one hands over the reasoning. Now the next attempt can be different in kind rather than merely in length.

This is the same principle as L3’s “what you already rejected”, applied in real time.

Three follow-ups worth having ready

• “What did you assume?” Surfaces the invented specifics before you build on them. Frequently the most valuable question in the whole conversation.
• “Give me three versions that differ in approach, not wording.” Stops you converging on your first idea. Genuinely different options, not synonyms.
• “What’s the strongest objection to this?” Useful precisely because the default register is agreeable. You have to ask for the pushback.

Know when to restart

Iteration has a failure mode. A long thread accumulates every rejected draft, and all of it is still in the context window shaping what comes next. Past a certain point you’re not steering, you’re dragging.

When a conversation has drifted, open a fresh one and write a single clean prompt containing everything you learned. Faster than a dozen more nudges, and usually better.

Try it now (6 minutes)

Take any answer you got in an earlier lesson and run three follow-ups: what did you assume, then a why-grounded correction, then what’s the strongest objection to this.

Notice that you learned more in three replies than in three fresh prompts.

Check your understanding

1. The first answer is most usefully read as:

A verdict on whether the tool is any good Evidence of how your request was actually read The best it can do without a paid plan

2. Replying only 'no, try again' usually produces:

A meaningfully different approach A variation on the thing you just rejected A request for clarification

3. A long thread full of rejected drafts is worth abandoning because:

Threads expire after a set length Every rejected attempt is still in context, shaping what comes next Restarting resets you to a better model

Check answers

Recap

The first answer diagnoses your request; follow-ups should carry the reasoning, not just the correction; and a drifted thread is worth abandoning for a clean one. Next: what to do when the answer is confidently wrong.