06-reference

every ai style guide

2026-03-19·reference·source: Every·by Katie Parrott
ai-writingstyle-guidesvoicepromptingcontent-craftevery-greatest-hits

AI Style Guides: How to Help AI Write Like You - Katie Parrott

Re-read at full length 2026-10-03 (paid).

Why this is in the vault

The canonical Every method for turning a writer's instincts into a reusable instruction set (seven sections, three maturity levels, a five-step interview build), which is the exact artifact The Denominator needs before Ray drafts anything in Ben's name. The earlier version of this note (~200 words) was built from a teaser and summarized a different framing (a history of style guides from Oxford compositors to AP); the full guide is a practical how-to with templates, and this rewrite replaces that summary.

The core argument

Parrott's diagnosis: models default to their "safest, most average approximation" of good writing, so loose prompts return canned transitions, puffery, and "not X, but Y" constructions. A traditional style guide solves inconsistency among many humans. An AI style guide solves the opposite: the model is already perfectly consistent, which is why it sounds like nobody. The guide's job is to push it toward one writer's idiosyncrasies. Her key distinction: a prompt says what to do; a style guide is the reusable layer underneath that says how to sound while doing it.

The seven sections (with Every's two house examples)

Every runs two working guides, one for Working Overtime (single-author column on tech and work) and one for Source Code (multi-author, building software with AI).

  1. Voice and tone - tensions and boundaries, not flattering adjectives. Working Overtime uses five rules (write "from inside the transformation, not above it"; participant plus observer; vulnerability as analytical). Source Code instead defines a spectrum of valid registers (technical authority, confessional practitioner, workshop narrator), which suits multiple authors.
  2. Structure - Working Overtime's arc: open with friction, document the messy experiment, surface the pattern, deliver a usable takeaway. Source Code keeps seven named templates (Discovery Arc, Technical Deep-Dive, Workshop Report, Multi-Case-Study, Gateway Guide, Two-Part Series, Q&A), each with steps and an archive example.
  3. Sentence-level preferences - paired concrete-vs-vague examples ("$400/month replacing $400k/year" instead of "cost-effective"); Working Overtime adds embodied specificity at emotional moments.
  4. Signature moves - named moves such as Friction to Zoom Out to Framework, the Adventure Narrative, the Borrowed Lens. Source Code's "One Big Idea" rule: every piece must coin, claim, or crystallize one memorable concept.
  5. Anti-patterns - called "often the highest-value section." A red-flag table of pattern and fix (hedges: delete; correlatives: rewrite; filler rhetorical questions: cut; more than three paragraphs before stakes: start with friction; recap endings: extend or reframe). Working Overtime also cites a data-backed killer: abstract industry analysis without personal stakes produced a 78% engagement drop. Source Code pre-answers named reader objections.
  6. Positive and negative examples - rules paired with examples beat rules alone.
  7. Revision checklist - built from the specific ways your drafts go wrong ("Stakes clear by paragraph one?", "Ending extends the idea, not summarizes?").

The copyable template splits examples into positive and negative, giving eight slots.

Three levels of guide

The five-step build

People react better than they self-describe, so have the AI interview you: (1) feed it 3-5 samples (or admired writing); (2) have it interview you one question at a time; (3) react to concrete sample lines ("too symmetrical," "too pleased with itself"); (4) synthesize into a draft guide; (5) use it for a week, then revise. A copyable starter interview prompt is included.

What makes it work

Be more specific than feels natural; spend effort on the blacklist (Working Overtime's started at three items and now has dozens, sorted by severity, each tied to a real draft failure); use examples; scope the guide to the writing you actually do (Source Code keeps separate guidance for engineering, practice, and philosophy pieces); don't make it rigid; and don't expect it to replace taste.

Mapping against Ray Data Co

The most direct hit is [[2026-10-02-site-spec]]: The Denominator has a locked masthead and draft posts ([[2026-10-02-agents-need-your-definitions-v1]]) but no publication-scoped style guide. RDCO has the inputs scattered across three places: the Sanity Check voice profile ([[voice-profile-2026-05-08]]), the house writing standard ([[WRITING-rdco]]), and the draft-review skill. By Parrott's ladder that is a Level 1-2 setup for a different publication. The Denominator is a multi-post analytical series about enterprise agents, which argues for her Source Code shape: named structural templates plus a "One Big Idea" rule, not a single-voice column guide.

Second connection: her Level 3 "compound system" is what [[2026-06-08-every-guardrails-review-skills]] already showed in practice (Parrott's /guardrails skill), and what RDCO's critic pattern does for vault notes. The gap is the feedback loop: RDCO checks drafts against voice but does not systematically write each correction back into the blacklist.

Third: the guide's anti-pattern table overlaps the em-dash and weasel-word rules in WRITING-rdco, and the "78% engagement drop" finding backs the existing no-derivative rule for Sanity Check: analysis without stakes is the failure mode.

Why it matters for RDCO / The Denominator

⚠️ Sponsorship

House promo, no third-party sponsor. The guide positions Spiral, Every's AI writing product, as built on this method, and the Level 3 "compound system" leads naturally into Every's Compound Writing plugin. Bias implication: the guide favors a tooled, plugin-shaped end state; the Level 1 and 2 advice stands on its own without any Every product.

Related