My AI writing style guide is not a taste manifesto. It is a failure log converted into enforceable sentences the next brief must satisfy. This post walks how config/style-guide.md grew from gate failures — with a short version timeline — inside my Cursor workflow and the AI writing pipeline that splits editor and writer.
The loop: fail the gate, sharpen the guide
The contract in AGENTS.md is simple: when a draft fails the gate, I do not hand-polish body prose in the expensive editor model. I fix the brief or add a style-guide rule, then re-delegate the writer.
That loop is what the brief gate checklist exists to enforce. Each miss that would otherwise become a habitual rewrite gets a sentence the next brief must meet.
The point of the split is cost and consistency. The editor invests in briefs, facts, and gates; the writer turns a brief into prose.
If the editor starts rewriting sentences by hand, the harness stops learning.
So a fail becomes one of two edits:
- Sharpen the brief so the next cheap writer cannot miss the fact
- Add a style-guide rule so every future brief inherits the constraint
The guide is the accumulated product of that loop. Cheap writers pass when the rules are sharp enough. Expensive rewrites mean the harness still has a gap.
AI writing style guide versions (what each scar bought)
Each version of my AI writing style guide locked in one scar, not a wishlist. Current label is harness_version: 6 as of 2026-07. These numbers are harness labels in this repo, not a public product scheme.
Early: readability and images
Early drafts piled into walls of AI prose that nobody wanted to skim. The first lock-ins were short paragraphs, lists for three or more items, image placeholders for the WordPress contract, and a basic voice split between manual and synth tracks.
Those rules sound basic until you watch an unguided model write a single block that never breaks. Readability arrived as a gate scar, not as design taste.
SEO placement rules
Focus keyword had to land in title, slug, and first paragraph, plus at least one outbound official or primary link. A branding nickname once stole the keyword from title and URL — the Log For AI lesson — so SEO placement became a hard rule instead of a hope.
The lesson was mechanical. If the SEO phrase is missing from title or slug, later checks fail no matter how careful the body is.
v3: publisher checklist
Body copy could look fine and Rank Math still collapsed to about 20. The focus keyword had never been set on the post.
That miss is easy to miss in the writing step. The writer can place the phrase correctly in Markdown and still hand the publisher a draft that fails every keyword check in WordPress.
The publisher checklist locked Rank Math focus keyword and meta description after the WordPress draft. The expected score band is about 75–85; I do not chase 100.
Attribution budget and evergreen voice
Early evergreen synth drafts spammed attribution hedges. The guide capped “public comparisons / third-party summaries” style phrases at five per article.
N:1 synth stays non-experience. When template leftover junk left the wrong persona or FAQ scaffolding, the fix was the brief and a re-delegate — not a hand polish.
That pair of rules keeps evergreen posts from sounding like either a press release full of hedges or a fake hands-on review.
v4: Commentary track
Hot takes needed a different contract than evergreen. v4 added the Commentary track: voice A or B, an as-of date, a primary source, and no Rank Math 75–85 chase.
Without that track, timely posts kept getting graded like evergreen tool comparisons. Freshness and a clear stance matter more there than density theater.
v5: rule 10 (2026-07-20)
Comparison drafts hedged pricing with lines like “commonly reported around $X” even when official pages stated a number. Rule 10 requires concrete official facts, an as-of date, and verified numbers supplied in the brief.
The editor contract is the real fix. If the fact sheet lacks a number the article needs, I fix the brief. I do not ask the writer to approximate.
Four posts were re-briefed and rewritten to de-hedge. Log notes show hedge counts dropped sharply on one pass — for example 15 down to 1.
v6: Methodology track (2026-07-21)
Experience posts needed a contract so writers stop inventing scars. v6 added the Methodology track: artifact citation, a human material contract before writer delegation, and at least one before/after for experience posts.
Without that track, methodology drafts collapsed into generic advice. With it, claims about rules and pipelines must point at real files, and personal scars wait until a human fills them in.
A rule looks like rule 10 in short form: state concrete facts with an as-of date, and make the editor supply verified numbers in the brief so the writer never approximates. I do not reprint the whole readability list here; one sharp rule shows the shape of the rest.
Two scars worth keeping
Two failures earned permanent keep because they kept repeating until the guide named them. Both show the same pattern: the body can look acceptable while the system still fails.
Failure A — Rank Math around 20
Before the publisher checklist, the keyword already sat in the body and the score still collapsed. Focus keyword was unset on WordPress, so almost every keyword check failed.
After the publish script and checklist set Rank Math focus keyword plus description, the intentional band is about 75–85. Remaining deductions are usually ignores I refuse to chase.
That scar taught me that “writing quality” and “publisher checklist” are different layers. Fixing prose cannot repair an unset focus keyword field.
Failure B — Hedged prices (v5)
Before rule 10, drafts said “commonly reported around $X” even when official pages stated a price. After the rule and fact-sheet contract, four posts were de-hedged.
The sharp version of the rule is boring on purpose. If an official page states a price or limit, write that fact with an as-of date. Hedge only when the claim is genuinely contested or missing from the brief.
That scar also matches how Google frames helpful content and experience: concrete, checkable claims beat vague hedges. I treat Google’s guidance on creating helpful content as a framing check, not a ranking promise.
What I refuse to add
A style guide does not replace taste or research. Too many rules become unread walls — the same failure mode as an overlong AGENTS.md. I promote only repeat misses.
I refuse to add:
- One-off taste nits that will not fire on the next draft
- Rank Math power words or chase-100 density games
- Rules that duplicate a skill’s entire procedure
Cursor agent rules that stick cover what belongs in agent rules. The style guide accumulates editorial scars the next brief must satisfy — different job, same discipline.
If a miss happens once, I note it and move on. If the same miss survives another gate, it earns a sentence in config/style-guide.md. That promotion rule is the only growth policy I trust.
Wrap-up
An AI writing style guide earns keep when the next cheap writer passes without expensive rewrites. Gate failures become sentences; sentences become the harness.
That is the methodology arc for this cluster: rules that stick, a pipeline that splits editor and writer, and a guide that only grows when the same miss shows up again. The file keeps evolving when the gate proves a new scar — not on a calendar promise.