Guide · Agent skill
How to use the Paper Writing Pipeline agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The paper-writing skill is an agent skill for Claude Code or another SKILL.md-compatible agent. It runs a multi-stage pipeline that starts from a research narrative report and produces a LaTeX paper and compiled PDF, with reviewer-driven revision rounds along the way. This guide explains what the skill does, how to install it, how to trigger it with natural requests, and what it needs, so you can run it on your own experiment results and check the output carefully.
Paper Writing Pipelinewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download and unzip the skill. Download the ZIP from our page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ in your project.
- 2Prepare your narrative report. Write NARRATIVE_REPORT.md with your claims, experiments, results and figure descriptions. The more quantitative detail, the better the draft.
- 3Start the pipeline with a request. Ask the agent to run the paper writing pipeline on your report, for example by saying 'write paper pipeline' with the file path.
- 4Confirm the assurance level. Check the announced assurance level and set it explicitly if you need submission-level audits.
- 5Approve the plan and contract. Review the plan summary and PAPER_ACCEPTANCE_CONTRACT.md. Set AUTO_PROCEED to false if you want to approve each phase.
- 6Review the compiled paper. Open the PDF, read the reviewer feedback and audit outputs, then verify facts, citations and numbers yourself.
What the paper-writing skill makes the agent do
The skill chains five sub-skills in a fixed order: paper-plan builds the outline, paper-figure generates plots, paper-write writes the LaTeX, paper-compile builds the PDF, and auto-paper-improvement-loop runs review-and-revise rounds. Each phase uses the previous phase's output.
The final deliverable is a paper/ directory containing the LaTeX source and the compiled PDF. Along the way the plan is checked against a Claims-Evidence Matrix, so every claim should map to evidence and every experiment should support a claim.
- Input: a NARRATIVE_REPORT.md file, a research direction with experiment results, or an existing PAPER_PLAN.md.
- Output: PAPER_PLAN.md, figures, LaTeX source, a compiled PDF, and reviewer feedback records.
- A reviewer model, reached through MCP servers, checks the plan, figures, writing and revisions.
Install the skill from the ZIP or the Claude apps
Download the ZIP for this skill from our page. The package contains SKILL.md. For a personal install, unzip it into ~/.claude/skills/. For a project-only install, unzip it into the .claude/skills/ folder inside your project.
If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.
- Personal: ~/.claude/skills/
- Project: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills → upload
Trigger the skill with a natural request
You do not need a special command to start the pipeline. The skill's description lists phrases that should activate it, including 'write paper pipeline', 'paper writing', '写论文全流程', and '从报告到PDF'. You can also name the skill directly with its argument hint, which expects a narrative report path or a topic.
Give the agent the path to your narrative report and say what you want, for example: 'Use the paper writing pipeline on NARRATIVE_REPORT.md and produce a full draft.'
- 'Run the paper writing pipeline on NARRATIVE_REPORT.md.'
- 'Write paper pipeline for my experiments in results/.'
- '从报告到PDF, using NARRATIVE_REPORT.md.'
/paper-writing "NARRATIVE_REPORT.md" — venue: NeurIPS, illustration: gemini, human checkpoint: trueSet options: venue, rounds, checkpoints and illustrations
The skill has defaults you can override inline. VENUE defaults to ICLR and can be set to NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL or IEEE_CONF. MAX_IMPROVEMENT_ROUNDS defaults to 2. AUTO_PROCEED defaults to true, so phases continue without pausing; set it to false to approve each phase. HUMAN_CHECKPOINT defaults to false; when true, the improvement loop pauses after each round so you can see the score and add instructions.
ILLUSTRATION controls the architecture or method diagrams. The default figurespec produces diagrams from JSON via /figure-spec. gemini generates AI illustrations and needs GEMINI_API_KEY. mermaid uses Mermaid syntax. false skips this phase so you draw diagrams manually.
- VENUE: ICLR (default), NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL, IEEE_CONF
- MAX_IMPROVEMENT_ROUNDS: 2 by default
- AUTO_PROCEED: true by default
- HUMAN_CHECKPOINT: false by default
- ILLUSTRATION: figurespec (default), gemini, mermaid or false
Use assurance levels to decide how strict the audits are
Phase 0 sets an assurance level, either draft or submission. You can pass it explicitly as '— assurance: draft' or '— assurance: submission'. If you do not, the level follows the effort setting: lite and balanced map to draft, while max and beast map to submission. The default is draft.
At submission level, four mandatory audits (proof-checker, paper-claim-audit, citation-audit and kill-argument) must each produce a JSON artifact and cannot be silently skipped. A verifier script, verify_paper_audits.sh, then gates the final report. The pipeline reports either 'Submission-ready: provisional' or 'Submission-ready: yes'. Use yes only when the verifier confirms coverage, including completed positive reviews from different model families where semantic review is required.
- draft: audits run only when their content matches; missing artifacts do not block
- submission: audits are required and verify_paper_audits.sh can block the final report
- provisional: same-family semantic audits are not fully independent
echo "<resolved-level>" > paper/.aris/assurance.txtAgree the acceptance contract before writing
Before any section is written, the pipeline drafts PAPER_ACCEPTANCE_CONTRACT.md with 10 to 20 testable assertions. Each headline claim needs a named evidence source, each abstract number must trace to a results file, and the title and abstract must carry the scope qualifiers the evidence supports. The contract is negotiated adversarially and graded at the end.
If you resume from an existing PAPER_PLAN.md, the contract step still runs. The source describes only one exception, a genuine legacy run from before this phase.
Check the output before you submit
The pipeline is designed to produce a reviewed draft, not a paper that is ready to submit unchanged. The skill's own notes say that same-family reviews are provisional and that fabricated or weak claims can still get through. Check every fact, citation and reported result against your data and sources.
Read PAPER_PLAN.md and the audit artifacts, then open the PDF and compare each number in the abstract with your results files.
- Verify each citation exists and supports the sentence it is attached to.
- Compare every abstract and results number with your experiment files.
- Confirm the scope in the title and abstract matches what you actually tested.
What people use it for
- Turning a finished set of experiments into a first full LaTeX draft for a conference submission.
- Producing a structured draft for a student thesis chapter or journal manuscript, targeting a venue such as IEEE_JOURNAL.
- Creating a reviewed paper draft with an audit trail that a research team can revisit and reproduce.
- Testing how a narrative report reads as a paper before deciding what experiments to add.
Frequently asked questions
Is the paper-writing skill free to use?
The skill is published under the MIT licence, as stated on the item page. Running it can still involve costs outside the skill, such as the agent platform and any reviewer model services it calls, and the source does not list prices.
What do I need to install and run the paper-writing skill?
You need Claude Code or a SKILL.md-compatible agent with Bash, Read, Write, Edit, Grep, Glob and Skill tools. The pipeline also needs the sub-skills paper-plan, paper-figure, paper-write, paper-compile and auto-paper-improvement-loop from the same repository, a reviewer backend through MCP servers, and a LaTeX toolchain to compile the PDF.
Does the paper-writing skill run code and use the network?
Yes. The skill runs scripts and uses the network, because the pipeline compiles LaTeX and calls reviewer model services. If you do not want code execution or external reviewer API calls, do not use this skill.
Do I need GEMINI_API_KEY?
Only if you choose the gemini illustration option. The default figurespec option and the mermaid option do not need that key, according to the skill's settings.
Can the skill write a paper without my research results?
It should not. The skill needs a narrative report or experiment data, and without them it will invent content. Provide your results first.
How does this compare with other paper-writing tools?
The source does not compare this skill with other tools. Its distinguishing features are the chained sub-skills, the reviewer-model improvement loop, and the assurance gates that label a draft as provisional or submission-ready.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you already have experiment results and want a structured first LaTeX draft with reviewer feedback and an audit trail, and you are comfortable checking every claim and citation yourself. Pick something else if you have no results yet, do not want the agent to run code or call external reviewer services, or need a paper that can be submitted without human verification.
Download the skill