The Three Drafts pattern

Most AI writing problems aren't quality problems. They're convergence problems.

You ask the AI to write something. It produces a draft. The draft is fine — competent, structured, legible. But it sounds like it was written by an AI. So you ask for revisions, and the revisions sound like AI revising AI. The output keeps drifting toward the same statistical center.

The pattern

Don't ask for one draft. Ask for three, with explicit instructions that they should be as different from each other as possible.

Write three completely different drafts of [the thing].

Draft 1: [tone/structure/angle 1]
Draft 2: [different tone/structure/angle]
Draft 3: [a third angle that doesn't overlap with the first two]

The three drafts should feel like they were written by three different
people with three different opinions about the topic. Don't average them.
Don't hedge. Lean into the difference.

Why it works

The model's first draft is its mode — the most statistically likely output for your prompt. The second draft is also pretty close to the mode. Asking for three drafts with explicit divergence forces it to sample from less-traveled paths.

You then iterate on whichever has the strongest spine. The result reads less AI-generated because it isn't the AI-generated default.

Example output (cold opener for a sales email)

Draft 1 — straight value: "Hi [name] — saw you just hired an AE. We help teams like yours close 30% faster. Worth a 5-minute reply about how?"

Draft 2 — observation-first: "That post about your win-rate problem — I've been seeing the same pattern across 12 teams this quarter. Have a quick read share if useful."

Draft 3 — provocative: "You're spending more on outbound than your top reps' base salaries. That's the wrong allocation. Here's what we'd change."

Three different people, three different bets. Pick the one that fits the situation.

What not to do

• Don't ask for "three variations." Variations end up rephrasing the same idea. Ask for different angles.
• Don't combine them. The averaged result is back to the mode.
• Don't use this for short prompts where the answer is unique. It's for writing, not factual lookups.

When it really shines

Email subject lines. Cold opens. Headlines. Tweets. Product descriptions. Anything where the angle matters more than the information.