06-reference

every compound writing guide

2026-09-02·reference·source: Every·by Katie Parrott
compound-writingai-writingclaude-code-pluginsclaude-skillsvoiceeditorial-processevery-greatest-hits

Compound Writing: The Ultimate Guide - Katie Parrott

Why this is in the vault

The full operating manual behind Every's Compound Writing plugin: a six-stage loop, a 30+ skill inventory, the VOICE.md / STYLE.md file architecture, and a five-step method for saving lessons. It is the most complete public template for the writing system The Denominator needs. The companion launch essay is already filed as [[2026-09-02-every-compound-writing-plugin]]; this note covers the guide itself.

The core argument

Parrott's problem was repetition: each piece meant re-explaining what she wanted and re-correcting the same output. Compound Writing makes each session start smarter than the last. It is three things: a methodology (idea to publication with review and learning built in), a toolkit (a plugin of reusable skills for Claude Code or Codex), and a "writing gym" where you choose what to push on. It is explicitly not a "vending machine": expect extensive back-and-forth.

The sharpest framing is how it differs from Compound Engineering. Code has deterministic checks (a test passes or fails), so a well-specified task can be delegated and reviewed later. Writing has some factual checks, but whether an argument convinces or a sentence sounds like you is qualitative; two editors can agree a paragraph is accurate and disagree whether it works. So Compound Writing asks the writer to "stay close to the work" rather than walk away.

The loop

Ideate/Interview → Outline → Draft → Review → Finalize → Compound. Each stage leaves an inspectable artifact (notes, outline, prose, critique, readiness decision, saved lesson), and you can enter at any stage.

Named persona reviewers add lenses: cw-asshole (least-charitable read), cw-hemingway (economy), cw-hitchcock (suspense), cw-mom (supportive general reader who gets lost), cw-sedaris (humor), cw-sorkin (pacing), cw-vonnegut (wants and stakes). cw-scribe is the router that picks the smallest useful workflow; cw-emergent combines skills for jobs none covers.

The file architecture

Plugin = recipe book, agent = cook, writing folder = kitchen. The plugin sets up VOICE.md (how sentences sound: diction, rhythm, humor, with good and rejected examples), STYLE.md (project standards: reader, purpose, argument shapes, evidence rules, publication requirements), examples/, and drafts/<piece>/ holding notes, outline, and versions. An optional AGENTS.md routes a desktop agent to the right project. Multiple publications can each carry their own STYLE.md with a shared VOICE.md. When inferring voice from samples, correct the agent: it may mistake one essay's subject for a permanent interest or promote a tic into a trait.

How to compound (five steps)

  1. Notice what you kept correcting or what unstuck you.
  2. Turn it into a specific rule with a why, plus a short before/after (her example: when tightening a sentence with a comic payoff, preserve the setup details).
  3. Decide where it belongs: VOICE.md, STYLE.md, a workflow checklist, or only this piece's notes.
  4. Save it with cw-save and confirm the file actually changed.
  5. Test it on the next piece; narrow rules that overfire and delete ones that stopped helping. Not every session needs a new rule.

Guardrails

Own the argument (you must be able to defend it unaided); verify facts against originals, keeping your experience, a source's findings, and the agent's inferences distinct ("An AI reviewer's approval is advice, not verification"); preserve attribution and never present generated quotes as reporting; do the selection work before asking anyone to read.

Mapping against Ray Data Co

The tightest connection is to The Denominator ([[2026-10-02-site-spec]]): Ben has a locked masthead and three post drafts in 01-projects/personal-brand/drafts/ ([[2026-10-02-agents-need-your-definitions-v1]]) but no publication-level STYLE.md. RDCO's current writing assets map onto Parrott's architecture unevenly: [[voice-profile-2026-05-08]] is a VOICE.md written for Sanity Check, [[WRITING-rdco]] is a cross-project standard, and draft-review plays cw-dev-edit plus cw-voice-check. Missing pieces: a Denominator STYLE.md, an interview-first entry point, and any cw-save equivalent. That last gap was already flagged in the launch-post note and is still open.

Second: her code-vs-writing distinction is the same one the phData copilot agent factory faces. Deterministic checks let you delegate and walk away; qualitative judgment does not. That is the honest boundary for how far a client's agent pipeline can climb before review becomes the constraint.

Third: Ray's vault pipeline already runs Finalize-style layers (schema audit, fresh-eyes critic). Her cw-tracks idea, a check for leftover process narration, is a useful addition for agent-written posts, where "I checked X, then Y" scaffolding leaks into prose.

Why it matters for RDCO / The Denominator

⚠️ Sponsorship

House promo, no third-party sponsor. The guide is distribution for Every's own open-source Compound Writing plugin and links Every's AI Style Guides and ChatGPT for Knowledge Work guides. Bias implication: the plugin is presented as the natural endpoint; the method (loop, file architecture, five-step compounding) is portable and works without it. It is free and open source, so the commercial pull is toward Every subscriptions, not a paid tool.

Related