doc-coauthoring helps with proposals, technical specifications, and decision documents that require shared understanding. Its main job is to transfer context from your head into a document that works without you standing beside it. It is more than sentence polishing, and it does not automatically produce a Word file. Final DOCX formatting can follow after the content stabilizes.
Install and invoke
Send these two commands in the Claude Code chat input, one at a time. They are plugin commands, not shell commands for your operating system. The second installs the official collection containing doc-coauthoring.
/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills
In your project conversation, ask Claude to “use doc-coauthoring” and supply the task or files. The prompt below is an editable starting example. Installing the collection supplies instructions and resources; it does not supply your source material or guarantee that every runtime dependency is installed locally. If the skill is missing, check that its plugin is installed, enabled, and loaded in the current session. Do that before troubleshooting the requested output.
Begin with a decision and an audience
Give the agent a concrete writing objective:
Use doc-coauthoring to write a decision document about replacing internal search.
The readers are engineering leadership and operations.
They need to decide whether to fund a two-week prototype.
I will provide the current problems, three options, and cost estimates.
Find missing context before drafting. Mark unknown data rather than inventing test results.
The first stage establishes the document type, audience, desired impact, template, and constraints. You can provide rough notes rather than organizing everything first. Answer the follow-up questions about gaps, alternatives, and trade-offs. If a conclusion has no supporting evidence, keep that gap visible. When a shared document cannot be accessed, supplying the relevant excerpt is often more productive than repeatedly trying an unavailable integration.
The second stage works section by section. For a decision document, start with the uncertain proposal or trade-off, leaving the summary until later. The agent can suggest candidate points; you choose what to keep, combine, or remove before it drafts the section. Targeted revisions make changing decisions easier to track than rewriting the entire document every turn.
Test with a reader who lacks the conversation
When the draft is complete, give a fresh conversation only the document. Ask why the work matters now, which alternatives were considered, and what evidence will determine the next decision after two weeks. A host with subagents can perform this check directly; otherwise, open a separate conversation yourself. If the reader cannot answer, repair the document rather than explaining the missing context only in chat.
This tests comprehension, not factual accuracy. A confident summary does not verify costs, legal interpretations, or performance claims. Those still need supporting material and responsible reviewers. Deliver the document with unresolved decisions and useful source links, so future readers can distinguish settled choices from research questions.
For a short notice, you can explicitly skip the full three-stage workflow. Supply an existing team template early to avoid restructuring late. If images carry essential meaning, add captions or text explanations so readers working from extracted text can understand the argument too.
Turn missing information into specific follow-ups
During context gathering, attach an owner and source to important numbers and judgments: finance estimates the cost, an existing report provides usage, and performance may still need measurement. Ask the agent to convert vague placeholders into answerable questions such as “Who can confirm migration cost?” and “What result would stop the pilot?” That makes the draft actionable.
For reader testing, give the finished document to someone who did not help write it, or to a fresh session with no chat history. Ask them to restate the decision, evidence, and remaining risk, then identify unsupported conclusions. If the answer depends on your spoken explanation, improve the document. If the data truly does not exist, keep it as a decision condition instead of inventing it.
When to shorten the three stages
If the source material and template are already complete, ask the agent to confirm audience and purpose briefly, then draft while retaining the reader test. If the material is scattered, use context dumping and questions before requesting polished prose. For a time-sensitive decision, begin with a one-page account of current state, options, recommendation, and conditions to verify, then expand. In every path, the document must explain the choice on its own. Agreement reached only in chat is not part of the deliverable.
Treat the first questions as a step, not delivery
This workflow asks for context before structuring and writing a document. An independent single-turn evaluation found that runs invoking the skill stopped at questions without writing the requested file. For CI or one-shot scripts, arrange a way to answer follow-up questions. If the task must finish in one turn with a draft, do not rely on this question-driven skill. Use a regular drafting task with complete inputs, permitted assumptions, and an output path, then verify that the file was actually created.