2026-08-19-styleguide

2026-08-19 — writing styleguide §

The author asked (2026-08-18, before bed) for a thorough styleguide of his writing, distilled from all of his notes, to be followed whenever an agent drafts prose into his part of the vault. Agent-authored docs in Claude/ keep their own style.

What was done §

• Full sweep. A 36-agent workflow read all 771 notes (~1.6MB) in 35 register-grouped batches: daily journal, weekly journal, long-form journal, atomic notes, inbox, books, projects/misc. Zero files skipped. Each reader returned structured observations across 14 style dimensions, distinctive tics, verbatim micro-examples tagged public/private, and outliers.
• Corpus statistics computed separately (bullet-marker counts, dash and ellipsis counts, question density, emoji, abbreviation variants, per-register and per-year splits) to ground the qualitative claims.
• Draft written to writing-style.md.
• Adversarial verification, second workflow (10 agents):
  • Five refuters re-read 50 held-out notes (10 per register group) against the draft, hunting contradictions and missing patterns. All five returned “needs-fixes” with concrete evidence.
  • A privacy audit traced every quoted example to its source; five example snippets traced to publish: false notes and were replaced with public-sourced or invented equivalents. All quotes in the final doc trace to publish: true notes.
  • An imitation test: one agent wrote three fake notes using only the guide; a blind judge got them shuffled with three real notes and identified all six correctly, listing which tells were deliberate (clean English — by policy) and which were guide failures.
• Revision folded everything back in. The biggest changes:
  • Added “rule zero”: the note being extended and its siblings outrank the guide — match local bullet markers, italics markers, voice, and section vocabulary. The draft’s global ”- bullets in all new writing” was wrong for weeklies, which still use * in 2026.
  • The register table went from one shape per folder to the shapes that actually occur (four distinct Notes/ shapes; stub and developed project notes; era-scoped daily/weekly skeletons).
  • Softened false “never”s the refuters disproved (Title Case, headings, concluding endings, blockquote attribution, “Thus”).
  • Added the judge’s imitation tells as new bans: repeated antithesis scaffolds, idiom stuffing, takeaway labels, resolving after the closing question, uniform paragraph geometry, unhedged opinions, invented sources, self-referential pointers.

Decisions §

• Voice, not forgery. The guide instructs writing correct English inside the author’s shapes — never planting his typos or L2 slips, and never “fixing” his existing prose. Clean English is the one acceptable tell in a draft. This was a policy choice, made before verification, and the blind-judge result confirmed it is the only reliable tell left.
• publish: false on the styleguide, deviating from the Claude/ publish: true convention: a detailed style fingerprint is an imitation manual for the author’s voice, and it catalogues his uncorrected errors — publishing either is his call. Flagged for his review; only he flips it.
• Long-form journal register is described but not draftable. The guide says propose, never write, for Journal/ long-form.
• New standing habit recorded (from the same instruction): commit and push to master as part of finishing work; anything new that would publish needs his OK first unless it is clearly not private or carries publish: false. Vercel deploy wiring stays untouched.

Rejected §

• Imitating the author’s typos/L2 slips for authenticity — pastiche, and disrespectful. Rejected before drafting.
• Publishing the guide with examples drawn from private notes even when innocuous — replaced all five that the audit caught.
• A single canonical “atomic note” recipe — verification showed it over-fit the 2026 claim-note register and would misfire on TIL, study, and older idea notes.

Open / unverified §

• The guide’s era-scoped claims (current weekly template, 2026 note shape) will drift; re-derive them when the newest notes stop matching.
• The imitation test covered three registers (claim note, TIL, book note). Weekly/daily/project drafts were not imitation-tested; their sections rest on the sweep plus refuter evidence only.