Claude doc-coauthoring skill

The last step of this skill hands your draft to a Claude that has never seen your conversation and asks it the questions a real reader would. Where it answers wrong, that section goes back for another pass. doc-coauthoring is one of Anthropic's example skills, added to the anthropics/skills repository in December 2025. It kicks in when you mention a PRD, design doc, decision doc, RFC, proposal or spec, and offers its workflow. Say no and Claude writes freeform. For how skills load, see how Claude Skills work.

Stage 1: context

Claude starts with five questions: what kind of document, who reads it, what it should change, whether there's a template, and any other constraints. Then you dump everything you have, unsorted: background, team discussions, why the obvious alternatives were dropped, deadlines, stakeholder worries. With connectors on, it can read Slack, Teams, Google Drive or SharePoint directly.

Next come 5 to 10 numbered questions about the gaps, which you can answer in shorthand ("1: yes, 2: see #channel"). The stage ends when Claude can ask about edge cases and trade-offs without needing the basics explained.

Stage 2: one section at a time

It starts with the section that has the most unknowns and leaves the summary for last. Each section goes through the same loop:

1. Clarifying questions about what belongs in it.
2. A numbered brainstorm of 5 to 20 points.
3. You curate: "Keep 1, 4, 7, remove 3, combine 11 and 12."
4. Claude drafts the section and then edits it surgically from your notes.

The skill asks you to say what to change rather than edit the draft yourself, so Claude learns your style for the next section. After three rounds with no real change, it asks what could be cut. Near the end it rereads the whole document for contradictions, repetition and generic filler.

Stage 3: the reader test

In Claude Code, Claude predicts 5 to 10 questions a reader would ask and runs each through a subagent that sees only the document. Another subagent checks for ambiguity, false assumptions and contradictions. On Claude.ai there are no subagents, so Claude gives you the questions to paste into a fresh chat yourself. The document is ready when the fresh reader answers correctly and finds nothing new.

When to use it

Use it for documents other people will act on: a design doc, a decision record, a proposal going to people who weren't in the meeting. For a weekly update or newsletter, the internal-comms skill has a fixed short format. The skill drafts in a markdown file or an artifact; if the result has to be a .docx, finish with the Word skill.

How to install it

Claude Code:

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

Claude.ai: the anthropics/skills README says its example skills are already available on paid plans.

Example request

Use the doc-coauthoring workflow. I need a [design doc / decision doc / proposal] about [topic] for [audience]. After reading it, they should [decision or action]. I'll paste my notes next; ask your clarifying questions after that.

Limits

• It asks two rounds of questions before it drafts anything.
• Its folder carries no LICENSE.txt and its SKILL.md names no license.
• Anthropic labels its example skills as demonstrations and advises testing them before you rely on them.