@nateliason - Nat Eliason Builds

Open original

@nateliason on X Nat Eliason describes a system he built in one Claude Code session that drafts posts in his own voice, with Gemini 4 Argon writing and Claude acting as judge. After a few rounds of feedback he says he now changes only about 3% of the words in the drafts it produces.

  • Argon writes every draft, Claude judges it against sealed predictions of Nat's edits, and the percentage of predicted edits is the signal for when he can step out of the loop.
  • Rules learned from one or two edits are treated as hypotheses until they hold across several pieces, after the judge overshot twice.
  • Edit rates fell from about 7% on the first live pieces to 10% and then 3% on video-based batches, with several posts going out unchanged.

The article is a walkthrough of how one writer built an AI pipeline that drafts pieces in his own voice, and it's more useful than most posts on the topic because it reports actual numbers and admits what failed. The core design is a split between two models. Gemini 4 Argon writes every draft, after the author ran a blind comparison against Claude and found Argon's prose preferable. Claude, running inside Claude Code, acts as judge: it scores each draft across eleven voice dimensions, checks every factual claim against the source notes, flags the worst lines, and feeds the top three fixes back to Argon for revision. Before the author ever reviews a draft, the judge writes down the edits it predicts he will make. Those predictions are sealed, then compared against his real edits, sorted into four buckets (voice rules, structure, invented facts, banned phrases). The share of his edits the judge anticipated is the metric that tells him when he can stop reviewing. His edit rate reportedly fell to 3%. Most of the practical value is in the preparation. The author cleaned roughly 44,000 words of his own published writing, stripping subscribe prompts, book plugs, and course footers, because otherwise the model learns to close every piece with an ad. He split the corpus by era, since his 2016 how-to posts sound younger than his current voice, and used the older posts only for structure. He wrote a voice guide from the training data before opening his holdout set, so the test pieces couldn't leak into the rules. And he checked every word on his ban list against his own usage, which surfaced a useful catch: words people treat as AI tells, like "foster" and "elevate," appear in his real writing, so banning them would have been wrong for him. The article also reports a failure worth noting. Some rules backfired, and the author documents them rather than presenting a clean success story. For anyone building something similar, the transferable lesson is less about the specific models and more about measuring against your own corpus instead of generic style advice.

Ask about this article