Doc Co-Authoring (Anthropic, official)
## what it does
tl;dr: a process skill, not a content skill. When the user says “write a doc / draft a PRD / make a decision doc”, the agent stops freewriting and runs a three-stage workflow — Context Gathering, Refinement & Structure, Reader Testing — with section-by-section brainstorming and a sub-agent reader test on Claude Code.
What it actually does
The SKILL.md we fetched on 2026-09-10 (from anthropics/skills/main/skills/doc-coauthoring/SKILL.md) is a 9.3 KB behavioral spec. The procedure in plain English:
Stage 1 — Context Gathering
Close the gap between what the user knows and what Claude knows. Ask five meta-questions (doc type, audience, desired impact, template, constraints), then encourage an info dump — channel links, shared docs, prior discussions, political context, timeline pressure. With 5–10 clarifying questions after the dump, the agent is allowed into Stage 2 only when “edge cases and trade-offs can be asked about without needing basics explained.”
Stage 2 — Refinement & Structure
Build the document section by section, in the order of most unknowns first. For each section: 5–10 clarifying questions, 5–20 brainstormed points, user picks what to keep/remove/combine, draft, then iterate with surgical edits. The skill tells the agent to use create_file for the initial scaffold and str_replace for every edit — never reprint the whole doc. Three consecutive iterations with no substantial changes triggers a “what can be removed” check.
Stage 3 — Reader Testing
Test the doc with a fresh Claude (no context bleed). On Claude Code the agent spawns sub-agents and asks the new Claude to answer 5–10 predicted reader questions; on claude.ai the user runs the test manually. Iterate on any gap Reader Claude surfaces. Exit when Reader Claude answers correctly and stops surfacing ambiguities.
Trigger conditions (from the frontmatter)
| Phrase | Routes in |
|---|---|
| ”write a doc”, “draft a proposal”, “create a spec”, “write up” | yes |
| ”PRD”, “design doc”, “decision doc”, “RFC” | yes |
| Any substantive writing task the user is starting | optional offer |
If the user declines the structured workflow, the skill says work freeform and stays out of the way.
Hard rules the skill encodes
| Rule | What it bans |
|---|---|
create_file for scaffold, str_replace for edits | Reprinting the whole doc on every change |
| Use artifacts only for drafting, not brainstorming lists | Using create_file for brainstorm dumps |
| Surgical edits only | Wholesale rewrites that lose user feedback history |
| Verify reader questions match what was actually written | Stale predictions carried forward from stage 2 |
| Three-stale-iteration check | Letting sections bloat past signal |
Install
The skill is one entry in the example-skills plugin of anthropic-agent-skills (per .claude-plugin/marketplace.json on 2026-09-10). Install requires the marketplace to be registered first.
# Claude Code — one-time marketplace registration
/plugin marketplace add anthropics/skills
# Install the example-skills plugin (includes doc-coauthoring)
/plugin install example-skills@anthropic-agent-skillsSibling skills in the same plugin: algorithmic-art, brand-guidelines, canvas-design, frontend-design, internal-comms, mcp-builder, skill-creator, slack-gif-creator, theme-factory, web-artifacts-builder, webapp-testing.
On Claude.ai, this skill is not pre-built — only the four document skills are. To use it in claude.ai you’d have to upload the folder as a custom skill (paid plan, Settings → Features → code execution on). On the API, this skill is instructions-only; there is no equivalent container.skills entry for it.
Gotchas
- Install command is
/plugin install, notnpx skills add. The README does not document annpxworkflow. Anyone runningnpx skills add anthropics/skillsis using a third-party installer, not Anthropic’s path. - Confined to
example-skills, notdocument-skills. A different plugin than PDF/docx/pptx/xlsx. Installingdocument-skillsalone will not get you this skill. - Slower than freeform. The workflow is intentionally heavier than “just write it.” If the user pushes back on the structure, the skill says to acknowledge the slowdown and offer a freeform path.
- Reader testing depends on the host. Sub-agent reader testing is a Claude Code feature. On claude.ai the user runs the test manually via a fresh conversation.
- Skill is instructions-only, not scripts. There are no companion
scripts/files; justSKILL.md. Eval vs prompt = pasteSKILL.mdinto context and run the same task twice. - Apache 2.0 for this one (per the README). It is the exception, not the rule — the four document skills in
document-skillsare proprietary. sandbox_testuntested. We read the SKILL.md and marketplace manifest, not exercised the workflow in a sandbox.
Related
- Anthropic PDF (official skill) — sibling in the document-skills plugin
- Anthropic DOCX (official skill) — sibling in the document-skills plugin
- Anthropic Academy Guide — same
anthropics/skillsrepo, but a routing skill over a live catalog - Claude skills guide — what the cache.directory skills shelf actually is
## compatibility
## embed this badge
