Doc coauthoring skill review: Anthropic's Doc Co-Authoring verdict
Doc Co-Authoring is Anthropic's one-file skill that interviews you and drafts a document section by section. A fresh reader then tests it. We rate it Conditional, 67/100, at Moderate confidence, because single-turn runs often produce no file.
- Verdict
- Conditional
- Score
- 67/100
- Confidence
- Moderate
- Safety grade
- A
- Best for
- People drafting a substantial spec, RFC, proposal or decision doc in an interactive Claude Code or claude.ai session
- Poor fit
- Headless, CI or single-turn pipelines that expect a finished file (#1718: 0 of 15 engaged runs wrote the file)
- Reviewed version
- anthropics/skills commit 8a1541c (2026-09-28); skill file unchanged since 0075614 (2025-12-04); no version
- Review date
- 2026-09-30
- Works in verified
- Claude Code
- Works in claimed
- Claude.ai
- Status
- no change since review (checked 2026-10-21); fix PR #1759 still open
Get the skill: doc-coauthoring folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing
What is the doc coauthoring skill?
The doc coauthoring skill is Anthropic's one-file workflow for writing specs, proposals and decision docs with Claude. It runs three stages: context gathering, refinement and structure, then reader testing by a fresh reader.
The skill is an Agent Skill, a SKILL.md file Claude loads when a request matches its description. See what Agent Skills are for the format. It turns scattered context into a PRD, design doc, RFC or proposal. SKILL.md names three stages:
- Context Gathering: Claude asks questions and reads shared docs and channels through connected integrations.
- Refinement & Structure: Claude builds each section from 5 to 20 brainstormed options you curate.
- Reader Testing: sub-agents receive only the document and a question. Claude.ai uses a manual fallback.
| Fact | Value |
|---|---|
| Publisher | Anthropic, in the anthropics/skills repository |
| Files | 1 SKILL.md, 15.8 KB, about 3,950 tokens |
| Scripts, references, assets | None |
| Added | 2025-12-04, commit 0075614 (PR #134) |
| Licence | No LICENSE.txt in the folder; no root licence in the repository |
| Bundle | Part of the example-skills plugin |
Several directories list the licence as Apache-2.0, but GitHub reports none.
Is the doc coauthoring skill worth installing?
Yes, the doc coauthoring skill is worth installing for interactive drafting only. It scores 67/100, Conditional, at Moderate confidence. Its reader-testing stage finds real gaps, but single-turn runs often end with no document.
It suits people drafting a substantial spec, RFC, proposal or decision doc in a live session. They answer several rounds of questions. It fits 3 cases poorly:
- Headless, CI or single-turn pipelines that expect a finished file on disk.
- Short documents, or users who want a one-shot draft.
- Non-Claude harnesses without create_file/str_replace tools or a fresh reader.
Stage 3 reader testing finds real defects. One hands-on report saw it flag a contradiction and six clarity problems. Nothing tells the caller that the skill needs several turns.
| Parameter | Score |
|---|---|
| Effectiveness | 72 |
| Reliability | 50 |
| Safety | 85 |
| Efficiency | 60 |
| Compatibility | 55 |
| Maintainability | 50 |
| Setup friction | 85 |
| Artifact delivery when interaction is impossible | 25 |
| Reader-testing value | 78 |
| Harness portability | 50 |
| Trigger precision | 55 |
| Instruction weight vs value | 65 |
| Cost of adoption | 80 |
Compare it with the best Claude writing skills, ranked by verdict.
How many tokens does the doc coauthoring skill use?
The doc coauthoring skill loads 3,954 tokens each time it triggers. A typical task costs far more.
| What | Tokens |
|---|---|
| Skill file, loaded each time the skill runs | 3,954 |
| One design doc, our estimate | 120,000 input, 15,000 output |
| One design doc, including the skill file | 124,954 input, 15,000 output |
The typical task is a 2,000-word design doc with reader testing, once a week. The skill stays in context across every question round. Stage 3 spawns 5 to 10 or more sub-agent calls, each carrying the full document (estimate).
Is the doc coauthoring skill safe to install?
Yes, the doc coauthoring skill is safe to install: it carries safety grade A. It is one Markdown file with no scripts, network calls or credentials. It writes and edits one document in place.
Our review of the file at commit 8a1541c found 8 facts:
| Area | Finding |
|---|---|
| Files | Single SKILL.md, about 3,950 tokens, last changed in the commit that added it (0075614) |
| Code | No scripts, references, assets, hooks, MCP config or environment variables |
| Writes | One markdown file in the working directory, or a claude.ai artifact, edited in place; no deletes |
| Reads | Shared docs and team channels through integrations; asks before searching connected tools |
| Untrusted content | No guidance on treating fetched channel or document content as untrusted data |
| Tool names | Claude.ai names create_file and str_replace, not generic write and edit verbs |
| Sub-agents | Stage 3 sub-agents get only the document and a question |
| Audits | No LICENSE.txt; skills.sh shows Gen Agent Trust Hub Pass, Socket Pass, Snyk Warn |
One directory shows "Grade A" at 100/100 for the ddmanyes copy, not this file. Our guide on how we grade skill safety explains the grades.
What does the community report about the doc coauthoring skill?
The community reports useful reader testing and a silent failure for the doc coauthoring skill. One hands-on report found a contradiction and six clarity issues. A 30-run test found no file whenever the skill engaged.
| Source | Date | Finding | Sentiment |
|---|---|---|---|
| anthropics/skills #1718 | 2026-09 | 30 Haiku 4.5 runs; all 15 engaged runs ended with questions and no file | Negative |
| anthropics/skills PR #1759 | 2026-09 | Direct-draft fix for non-interactive runs, still unmerged | Negative |
| Tessl quality review | 2026-09 | Quality 74%; workflow clarity 5/5; conciseness 3/5; duplicated branches | Neutral |
| koldovsky PR #25 | 2026-08 | Stage 3 found a contradiction (5 vs 3 test cases) and six other issues | Positive |
| koldovsky PR #17 | 2026-08 | Usable with minimal adaptation; only the trigger description was tightened | Positive |
| LesterThomas PR #4 | 2026-07 | A non-developer guide produced with the workflow, reader testing included | Positive |
| SkyScow #34 | 2026-09 | Stage 3 hard-codes a fresh Claude reader | Negative |
| SkyScow #98 | 2026-09 | create_file and str_replace do not exist in a write/edit/task harness | Negative |
| SKILL.md at 8a1541c | 2026-09 | Claude.ai tool names and no non-interactive path | Negative |
| anthropics/skills #1087 | 2026-05 | Installing document-skills loads this skill and 12 other undeclared skills | Negative |
PR #25 found an unreconciled reference, two undefined terms and a reproducibility gap too. Tessl, a skill quality registry, showed 74% at our review and 79% on its live page on 2026-10-20.
What are the doc coauthoring skill's known failures?
The doc coauthoring skill has three known failures at commit 8a1541c. Single-turn runs end with questions and no file. It names claude.ai-only tools, and its reader test assumes a fresh Claude instance.
The worst failure, listed first, ends a session with no document and no error.
- No document in single-turn or headless runs: 0 files in 15 engaged runs (#1718).
- Claude.ai tool names create_file and str_replace (SkyScow #98).
- Hard-coded fresh Claude reader in Stage 3 (SkyScow #34).
- Broad trigger: it fired 10/10 on a proposal prompt, and 5/10 once a file path was added.
- Duplicated artifact and file branches in about 370 inline lines (Tessl).
Which agents does the doc coauthoring skill work with?
The doc coauthoring skill is verified in Claude Code and claimed for Claude.ai. Directory claims that it works with any LLM lack evidence. Two reports show its tool names and reader break under GPT and DeepSeek harnesses.
| Agent or harness | Status | Evidence |
|---|---|---|
| Claude Code | Verified | koldovsky PR #17 and PR #25 |
| Claude.ai | Claimed by Anthropic | SKILL.md artifact branch; Anthropic README |
| GPT and DeepSeek harnesses | Fails when followed literally | SkyScow #34 and #98 |
| Codex | Unknown | No report found |
| TypingMind, mcpmarket listings | Claimed by the listings; no published run | Directory pages only |
See Claude Code skills ranked by verdict for skills verified in the same agent.
How do you install the doc coauthoring skill?
Install the doc coauthoring skill in Claude Code through the example-skills plugin from the anthropics/skills marketplace. In Claude.ai, Anthropic lists it among built-in example skills, with conflicting plan statements across its sources.
In Claude Code
- Add the marketplace:
/plugin marketplace add anthropics/skills. - Install the bundle:
/plugin install example-skills@anthropic-agent-skills. - Copy the folder to
~/.claude/skills/doc-coauthoring/instead, for this skill alone.
Installing the document-skills plugin loads it too (#1087). Our Claude Code skills setup guide covers scopes and folders.
In Claude.ai
Open Settings > Capabilities, or Customize > Skills, and switch the example skill on. Two Anthropic sources disagree on plans. The Help Center lists skills on Free, Pro, Max, Team and Enterprise. The anthropics/skills README says example skills are "already available to paid plans in Claude.ai". Our walkthrough shows how to turn on skills in Claude.
How do you use the doc coauthoring skill?
Use the doc coauthoring skill in a live session where you can answer questions. Ask Claude to draft a spec or proposal. Answer each round, curate the section options, then let the reader test find gaps.
- Start a session and ask for a design doc, proposal or decision doc.
- Answer the Stage 1 questions and point Claude at the shared docs.
- Curate the 5 to 20 options Claude offers for each section.
- Run Stage 3 and fix every gap the reader reports.
Evidence backs 3 uses: drafting proposals, reader-testing finished docs, and repository documentation. Unattended file generation fails in #1718.
Test it yourself:
- Run the skill once interactively on a real upcoming doc.
- Run it once with
claude -p, asking for a file. - Compare the reader-testing value with the no-output case.
What are the alternatives to the doc coauthoring skill?
The best alternative to the doc coauthoring skill depends on the job. Brainstorming, Recommended at 76, shapes ideas before writing. Internal Comms, Conditional at 64, drafts status updates, and Anthropic's DOCX skill produces Word files.
| Skill | Verdict and score | Better choice when |
|---|---|---|
| Brainstorming (obra), see the Brainstorming skill verdict | Recommended, 76 | You need a design gate and clarifying questions before code |
| Internal Comms (Anthropic), see our Internal Comms review | Conditional, 64 | You write 3P updates, newsletters or FAQ roundups in a house style |
| DOCX (Anthropic) | Not rated here | You need a Word file, tracked changes or comments |
| Writing Plans (obra) | Not rated here | You turn an approved spec into a test-first implementation plan |
Doc Co-Authoring decides what a document says; the DOCX skill decides the file format.
Where is the doc coauthoring skill on GitHub?
The doc coauthoring skill lives in the anthropics/skills GitHub repository at skills/doc-coauthoring. It is one SKILL.md, unchanged since 2025-12-04, with no licence file. GitHub code search finds 1,696 hits, the original and forks included.
- Repository: github.com/anthropics/skills, folder
skills/doc-coauthoring/. - Origin: commit 0075614 on 2025-12-04 (PR #134).
- File hash at review:
sha256:2e47d788…. - Licence: no LICENSE.txt in the folder.
- Installs: 87.7K on skills.sh, a skill install directory.
- Contributors: 15 in the repository.
Copies circulate under other names: claude-office-skills, claudiodearaujo on mcpmarket, ddmanyes and brycewang-stanford. Install from anthropics/skills, never a copy.
Common questions
What is a Claude document skill?
The document skills are DOCX, PDF, PPTX and XLSX, shipped in the document-skills plugin. Doc Co-Authoring is an example skill; see Anthropic's PDF skill and the PowerPoint skill review.
Can Codex use the doc coauthoring skill?
Not yet shown: no verified report shows Codex running it. Harness reports show its Claude tool names break outside Claude.
Is there a skill-creator skill in Claude?
Yes, Anthropic publishes one for building new skills. Read our skill-creator verdict before you use it.
Does the doc coauthoring skill write the file for me?
Yes, in interactive sessions only. Issue #1718 shows single-turn runs ending with questions and no file.