Doc Co-Authoring

Never start a doc from a blank page.

Most documents fail in the gap between what the author already knows and what the reader needs spelled out. Doc Co-Authoring runs an active three-stage workflow: it pulls your context out first, builds the doc one section at a time, then tests it on a fresh Claude with no memory of the conversation — surfacing the blind spots before a real reader hits them.

AnthropicPRDdecision docspec / RFCreader testing

Three stages

Context gathering  →  Refinement & structure  →  Reader testing

• Context gathering — you info-dump; Claude asks the clarifying questions until the gap closes.
• Refinement & structure — each section is brainstormed, curated, drafted, then edited surgically.
• Reader testing — a context-free Claude answers predicted reader questions; whatever it gets wrong is a gap to fix.

Key concept — reader testing

The standout idea: before anyone else reads the doc, a fresh Claude with zero context is asked the 5–10 questions a real reader would ask. Where it answers wrong, contradicts itself, or assumes missing knowledge, that’s a blind spot the authors couldn’t see — so the workflow loops back and fixes those sections. It mirrors how readers will actually consume the doc (often by pasting it into Claude).

Worked example — “help me write a decision doc”

1. Meta-questions — doc type, primary audience, desired impact, template, constraints.
2. Info dump — you brain-dump background, threads, trade-offs; Claude tracks gaps and asks 5–10 clarifying questions.
3. Scaffold — a file (or artifact) is created with every section header and [To be written] placeholders.
4. Per section — clarify → brainstorm 5–20 options → you keep/remove/combine → draft → refine.
5. Reader test — predicted questions go to a fresh Claude; gaps loop back to refinement. Done when it answers cleanly.

Under the hood

Stage 1 — context gathering

Opens with five meta-questions (type, audience, impact, template, constraints), then invites a stream-of-consciousness dump — channels to read, docs to link, politics and timeline pressure included. Connectors (Slack, Teams, Drive, SharePoint) pull context directly when available. Exit condition: Claude can ask about edge cases and trade-offs without needing the basics re-explained.

Stage 2 — the per-section loop

Start with the section that has the most unknowns (the core proposal for a decision doc, the technical approach for a spec); leave summaries for last.

1. Clarify — 5–10 questions on what the section must cover.
2. Brainstorm — 5–20 numbered options, including angles you may have forgotten.
3. Curate — “keep 1,4,7; remove 3 (dupes 1); combine 11+12.”
4. Gap check — anything important missing?
5. Draft + refine — edits are surgical (str_replace), never a full reprint.

After three quiet iterations it asks what can be cut without losing meaning.

Stage 3 — reader testing

• With sub-agents (Claude Code): each predicted question goes to a fresh sub-agent given only the document; results are summarised, plus checks for ambiguity, false assumptions and contradictions.
• Without sub-agents (claude.ai): the user runs the same test in a clean conversation and reports back what Reader Claude misread.

Either way, the doc is ready only when Reader Claude answers consistently and surfaces no new gaps.

Describe the change, don't make it

Through the section loop, you’re asked to say what to change (“tighten the third paragraph”, “drop the X bullet — Y covers it”) rather than editing the draft yourself. That’s how the skill learns your style and applies it to the sections still to come.

Read next

Internal Communications Short-form company updates.

Document processing Turn the finished doc into DOCX or PDF.

Anthropic Skills reference The full official catalogue.