Prose sensors for LLM-drafted writing

## Spark

Issue [#61 "Prose lints cleanly via Vale"](https://github.com/alunduil/blog.alunduil.com/issues/61) closed this week, with [#88 (UK spell-check via Hunspell en_GB)](https://github.com/alunduil/blog.alunduil.com/issues/88) and [#89 (per-path overrides for opinion vs formal voice)](https://github.com/alunduil/blog.alunduil.com/issues/89) queued behind it. The interesting bit isn't that I added Vale — it's *what I dropped to get clean lints*: FirstPerson, OxfordComma, write-good, Readability. Those rules were designed for human writers correcting human tics. The rules I kept catch what an LLM gets wrong drafting in my voice.

## Why it could be interesting

The published prose-linter advice almost universally lives at the human-writer layer: passive voice, weasel words, readability scores, complexity indices. With an LLM drafting in conversation, the failure modes are different: model-default em-dash patterns, Americanisms creeping into UK spelling, register drift (formal-ish output where I want opinion), generic transitional phrases ("Furthermore", "It's worth noting that").

The post-worthy claim: **for LLM-collaborated long-form writing, the rules you *remove* matter as much as the rules you keep**, and the kept set is shaped less by writing-quality dogma than by which model-tics show up in your specific voice. Distinct from #66 (lint-on-edit beats lint-on-PR — that's about *when* to run sensors) and #73 (Claude-native drafting flow — that's the whole pipeline). This is specifically: **what rule sets actually catch model-tics vs. human-tics, and what calibration looked like.**

## Open questions

- [ ] Catalog of model-tics actually caught vs. missed by the kept Vale style packages.
- [ ] Why did dropping FirstPerson/OxfordComma/write-good/Readability *help* (rather than hurt)?
- [ ] What does per-path voice (#89) actually buy — is "opinion" mode just "fewer rules", or does it need its own rule set?
- [ ] Does the catalog generalise beyond this blog to any LLM-collaborated long-form writing?

## Source material

- [blog#61](https://github.com/alunduil/blog.alunduil.com/issues/61) — Prose lints cleanly via Vale (closed)
- [blog#87](https://github.com/alunduil/blog.alunduil.com/pull/87) — Lint prose cleanly via Vale (PR)
- [blog#88](https://github.com/alunduil/blog.alunduil.com/issues/88) — Vale UK spell-check via Hunspell en_GB
- [blog#89](https://github.com/alunduil/blog.alunduil.com/issues/89) — Vale per-path overrides for opinion vs formal voice
- Commits in window: `dc206a2` Drop FirstPerson and OxfordComma overrides · `84de5cd` Tighten Vale config: drop write-good and Readability · `0a288c7` Unspaced em-dashes in scaffolding prose · `af1ad77` Wire Vale into pre-commit hooks · `d08f9f3` Add Vale config and synced style packages
- Related kernels: #73 (Claude-native drafting flow), #66 (lint-on-edit for agents)