Skip to content
Cameron Mills

All projects

A house style check that fails a rewrite which loses a number, name or citation

The humaniser and writing-style system

Part of Life Tracker

A scorer, a written house guide and a review that every agent report goes through, measured against my own writing, with a compare step that fails any rewrite that loses a number, name or citation.

Code on GitHub

Problem

Every agent in my system writes reports and messages for me. A day's worth of them showed the same habits again and again: long dashes, run-on lists and opening sentences that buried the point. I wanted one house style, checked by a script, and a rewriting step that could not quietly change what a report says.

Approach

The standard is my own measured writing, not "sound less like AI". It is a clarity and house-style check, not a tool for getting past AI detectors, which are unreliable in both directions. A score is evidence for an editor, never a verdict on who wrote something.

Architecture

A body of my own writing is measured into a profile for each kind of writing, and that profile passes through the same privacy check as other shared files. The writing itself is never shared, and nothing from it is quoted here. One guide holds the house rules, and every agent prompt that needs them reads them from it, so a change to the guide reaches the next run. Every managed agent run finishes with a style review of what it wrote. It then hands the report to a fresh editor agent, which reads it as the reader would, checks each fact against its sources and returns a verdict.

What I built

A single tool with four jobs. It scores a draft for the habits above, marks passages that must not be touched, and compares a rewrite with the original, failing if any number, name, citation or quotation has gone missing. Its review fails on any breach of the house rules, so a script can refuse a draft. A rewrite is saved beside the original, and the original is kept if anything was lost. It refuses to touch assessed work outright. Asking for a style review is all it takes to start it.

Results

In a test with its pass criteria fixed in advance, the guide cut style faults sharply where a plain "improve this" instruction made them worse, and kept length better. It still met only some of the criteria, so the verdict was "failed", and it stands.

What didn't work

The first measurement of my writing swept in material I had not written, such as other people's mail and the system's own emails sent from my account. Each is now filtered out. One stray punctuation mark in a short email once looked like a huge deviation, which was fixed with a minimum. The compare step first failed almost every rewrite by mistaking bullet labels for facts. It still does not check meaning: a misplaced "not" still needs a human reader.

Stack

  • Python
  • Claude Code
  • Markdown
  • LaTeX

How these figures are sourced

The outcome figure comes from an experiment: measured once and reported against criteria fixed before it ran. The status was read from the code repository at commit c8e90f7 on 1 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.