What we’ve tested

Do good samples need a written style guide?

What this shows: no clear overall benefit from adding this particular guide in this small pilot. The guide was preferred more often on voice alone, but not overall. Samples stay the practical default, and a guide stays optional. Milton stores your approved samples and corrections, and the assistant you already use does the writing.

Working conclusion: capable models may be able to follow a person's style from good examples, without an elaborate written guide. That is our hypothesis. We have not measured it.

What we measured: whether adding one particular guide to the same samples made first drafts better.

• Samples: three Paul Graham essays.
• Tasks: 10 fictional briefs, made up of 3 essays, 3 posts, 2 emails, and 2 DMs.
• Drafts: 20 first drafts from GPT-6.1-Sol at medium reasoning, two for each brief.
• Judging: Claude Opus 5.5 at high effort, working independently of the writer, compared each pair twice, once in each candidate order. It was not told which draft had the guide.
• Controls: within each pair, the samples, model, and brief were the same. Only the guide differed.
• Guide: the latest one supplied from Spiral, another writing tool. It was fixed before the drafts were written and not revised after we saw them.

A task counts for one side only when both orders chose that side. Mixed results count as inconclusive. All 20 drafts met their requested word ranges.

What this doesn't show:

• What people prefer. The judge was an AI, not a person, and we measured neither human preference nor editing time.
• Anything broader than this setup: one familiar author, one writer model, one guide, and one judge.
• How much samples help. Every draft had samples, so nothing was compared with drafts written without them.
• That all GPT or Claude models work equally well, or that Milton achieves any percentage of Spiral's quality.
• How a real person writes email or DMs. Those drafts carry a style over from essays, not from real correspondence.

Full notes: Method and results and All briefs and drafts.

Why Milton doesn't rewrite your writing

Milton stores your rules exactly as sent and your samples exactly as approved. Reading them back is a database query, with no model call. Milton itself never rewrites what you saved. Your assistant does the writing, with its own model. Milton keeps the file.

We tested adding a model to how Milton picks samples. See below.

How Milton picks eight samples

Milton keeps every sample you approve and serves eight per report. Your assistant says what it is writing, to whom, and why. Milton ranks your samples by who they were for and what they were for, then by matching words, then by date.

We tested that against two alternatives: the newest eight, and Google's embedding model, which matches by meaning. We ran it twice, on fictional writers, with drafts scored blind:

• Round 1: one writer, 150 samples, 20 drafted situations.
• Round 2: three writers, 1,500 samples, 120 drafted situations.

Embeddings found slightly better matches and wrote no better drafts. They would also send your samples to a third party on every read. We left them out.

The result we didn't expect: finding better samples did not produce better drafts. Our fictional writers wrote the same way to everyone, so any eight of their samples sounded alike. Real people write differently to a client than to a friend, which is where picking the right sample should matter. That test needs real writing, and it is next.

How many samples

Real writing teaches an assistant more than nothing. In one study, five real emails roughly doubled how often a model's email could be traced to the person it was imitating, from 29% to 60%, and two emails did about as well as ten. That is why setup asks for one email first, and why your samples outrank the setup choices.

In our earlier ranking tests, cutting from eight samples to four made drafts clearly worse, 3.21 against 3.65 out of 5. Milton serves eight.

What we haven't tested yet

• Drafts for real people, judged by those people.
• How well ChatGPT, Codex, and other clients beyond Claude Code retrieve samples from Milton and write from them. The style-guide pilot was an offline prompt test: the samples were placed in the prompt, and no client fetched them from Milton.
• Whether one short email is enough.
• Whether the setup choices, or rules a person has approved or corrected, add anything beyond a pasted sample. The style-guide pilot tested one fixed guide, not those.
• Whether you can pick out your own writing, blind.

How the earlier ranking tests were run

• The fictional writers and their samples were written by a model.
• Drafts were scored blind by a second model, twice each, against a description of the writer.
• No person has scored them yet. People are generous judges of their own voice: in one study, they rated drafts they had edited as just as much their own as text they wrote alone, though the drafts were measurably closer to the model's. So a person's score needs to be blind too.
• We set the pass mark before each run.

Full notes: Sample ranking and What the evidence supports.

Sources: Wang et al. (2025), Catch Me If You Can? Not Yet; Baumler et al. (2026), Can You Make It Sound Like You?.