Iterative refinement

The concept

"Make it better" is the least useful instruction in AI. The model has no idea what dimension of "better" you mean, so it makes something different — usually longer, more adjectival, and further from what you wanted. Getting good output is rarely a one-shot event; it's a short loop of surgical feedback. The skill is diagnosing which specific thing is wrong and naming it, exactly as you would in a creative review: you don't tell an art director "make it better," you say "the logo is fighting the headline, drop it to the bottom right and give me more air at the top."

Three rules for the loop:

1. Name the dimension, the direction, and the amount. Not "shorter" — "cut it by 40%, lose the second paragraph entirely." Not "warmer" — "warmer: use contractions, drop 'we would be delighted to,' address her by first name." Not "more creative" — "the idea is too safe: give me three versions where the central metaphor isn't a journey, a family, or a sunrise."

2. Keep what works, explicitly. The most common refinement failure is losing the good part while fixing the bad part. Say so: "Keep paragraph 1 exactly as-is. Rewrite only paragraph 2 to be more concrete. Do not touch the ending." Or quote the line you love back at it: "'This is not a threat, it's a schedule' — keep that energy, apply it to the other five."

3. Restart rather than pile up. After four or five rounds of edits, the chat is cluttered with rejected drafts and contradictory instructions, and quality degrades. When you're 80% there but circling, start a fresh chat: paste the current best version, state the remaining three problems, and ask for one clean pass. This "reset with the good draft" move rescues more stuck sessions than anything else.

The critique–revise loop. The advanced version — and the one that produces genuinely better work — is to make the model criticise its own output before revising. This works because generating and evaluating are different tasks, and models are noticeably better critics than they are first-draft writers.

Now switch roles. You are the client's CMO — sceptical, short on time,
has seen 40 agency decks this year.

Critique the draft above: list the 5 weakest points, in order of how
badly they'd damage us in the room. For each, say specifically what's
wrong, not just that it's weak.

Then rewrite, fixing all 5. Change nothing else.

Run that once and the output improves noticeably. Run it twice and you usually hit diminishing returns — a third pass tends to sand off the personality. Two passes, then your own hand on it.

Before → After

This is too generic, make it better and more creative and punchier.

Three specific problems, fix only these:

1. The opening line ("In today's fast-paced digital landscape") is
   dead copy. Cut it. Start on the tension: the client's customers
   don't trust delivery dates.

2. Section 3 makes a claim with no proof. Either attach a number
   from the brief or delete the claim.

3. The CTA is passive ("we look forward to discussing"). Make it a
   specific proposed action with a date.

KEEP: the structure, the three headings, and the closing paragraph —
those are right. Do not rewrite them.

Return the full document with only those three changes.