06-reference

writewithai cowork skill for fiction voice

2026-05-13·reference·source: WriteWithAI·by Cole Schafer & Dickie Bush
claude-skillsvoice-matchai-writingfat-skillswriting-voice

Why this is in the vault

Cole/Dickie walk through a concrete 6-step recipe for building a Claude (Cowork) .Skill file that captures an author's fiction-writing voice. The recipe is a layperson's articulation of the exact pattern RDCO uses internally: conversation-with-Claude to extract tacit knowledge, then crystallize into a reusable .Skill file referenced automatically by context.

Their phrasing matches the founder's own frame almost verbatim: "you can't automate what you can't articulate" - that's the same bottleneck RDCO addresses via fat-skills (per Tan). The first half is a thoughtful Sanderson-grounded piece on creative-atrophy risk; the second half is the technique. The whole thing reads as genuine technique not a Cowork product pitch (though it heavily promotes their own owned products: AI Writing Skool, Ship 30 for 30, the new Claude Code marketing bootcamp).

The 6-step process they describe:

  1. Open Claude desktop app (Cowork requires it)
  2. Ask Claude to walk you through a .Skill building exercise for your voice
  3. Ask Claude what additional context it needs (POV/tense, profanity, themes, emotional range, etc.)
  4. Provide at least 2 writing samples for pattern recognition
  5. Export the conversation as a .Skill
  6. Upload via settings → capabilities → Add, restart Claude

The "ask Claude what else it needs" loop is the cleverest part - they let the model surface gaps in the author's own articulation, which is more thorough than top-down prompt engineering.

Mapping against Ray Data Co

Strong mapping on two axes:

  1. Fat-skills pattern. "A .Skill is just context saved as a file (that AI can reference at any point in the future)" is a near-identical restatement of Tan's thin-harness/fat-skills thesis already in the vault at [[2026-04-11-garry-tan-thin-harness-fat-skills]]. WriteWithAI is independently arriving at the same conclusion from a writer's angle. Useful corroboration that the pattern is generalizing beyond engineering use cases.

  2. Voice-match technique transferability. RDCO's existing [[~/.claude/skills/voice-match/SKILL.md]] does Sanity Check voice consistency checking against the founder's published articles. The technique Cole describes - conversation-driven extraction of voice rules, then 2+ samples for pattern recognition - is a richer build pattern than the current voice-match skill, which is comparison-only against existing artifacts. Action candidate: run the founder through Cole's 6-step exercise to produce a founder-fiction-voice.Skill and a founder-newsletter-voice.Skill that voice-match could reference proactively (not just compare against post-hoc). Today voice-match is reactive scoring; the WWA recipe makes it generative-ready.

Bonus mapping: Cole's "ask AI what it needs" loop maps to RDCO's pipeline-critic seat that runs convergence loops - same idea of letting the model surface gaps in the human's articulation rather than relying on top-down spec. See [[concepts/2026-05-12-rdco-pipeline-rlhf-shaped]] for the parallel.

Skip the framing pieces: the Sanderson/atrophy preamble is good writing but adds nothing the vault doesn't already have via the [[feedback_calibrate_overconfidence]] and IC-vs-production-mode discipline notes. The Cowork product positioning ("No-Code Claude Code", desktop-app-required) is mostly marketing - the underlying mechanism is the same Claude Code skill system RDCO already uses.

Related