Guide · Agent skill
How to use the Paper Plan Outliner agent skill (paper-plan)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The paper-plan skill is an instruction package for an AI coding agent such as Claude Code. When you ask the agent to plan a research paper, it reads your project files, collects your claims, evidence and reviewer feedback, and produces a section-by-section outline sized to a target venue's page limit. This guide explains what the skill does, how to install it, how to trigger it, and what you need before you start, so you can get a usable paper plan before you begin drafting.
Paper Plan Outlinerwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Download the paper-plan skill ZIP from this page.
- 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use, or into your project's .claude/skills/ folder.
- 3Or upload in the Claude apps. In the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.
- 4Gather project files. Place NARRATIVE_REPORT.md or STORY.md, review-stage/AUTO_REVIEW.md and your experiment results in the project.
- 5Ask for a paper plan. Request a paper outline using a trigger phrase such as "paper outline" or "plan the paper", and name the venue.
- 6Review the outline. Check the Claims-Evidence Matrix, section lengths and figure plan against your data before drafting.
What is an agent skill, and what does paper-plan make the agent do?
An agent skill is a folder containing a SKILL.md file with instructions that an AI agent loads when a request matches the skill's description. The paper-plan skill is defined by a SKILL.md file with the name paper-plan. Its description says it generates a structured paper outline from review conclusions and experiment results.
When it runs, the skill has the agent gather core claims (usually 3-5), the evidence for each claim, known weaknesses from reviewer feedback, and suggested framing. It builds a Claims-Evidence Matrix with columns for claim, evidence, status and section. It then plans each section, covering length, key citations, the hero figure and the figures and tables to include, and finishes with a figure plan table.
- Reads narrative documents such as NARRATIVE_REPORT.md or STORY.md, and review-stage/AUTO_REVIEW.md
- Uses CLAIMS_FROM_RESULTS.md first if it exists, since it is the preferred claim source
- Chooses a paper structure (empirical, theory plus experiments, or method) and checks the page budget against the venue
- Lists figures with type, description, data source and priority
Install the paper-plan skill
Download the ZIP for the skill from this page and unzip it. For a personal install, place the unzipped skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.
If you are a Codex user who wants Claude Code to act as the reviewer, the source says to install this package after skills/skills-codex/*. The reviewer in that setup is a different model family from the Codex executor.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project
- Claude apps: Settings, then Capabilities, then Skills, then upload
Prepare your project files
The skill works from files already in your project. It expects one or more of the following. If none exist, the skill asks you to describe the paper's contribution in 3-5 sentences, but the plan will be weaker.
The skill also refers to a shared-references folder containing writing-principles.md, venue-checklists.md and assurance-contract.md. It reads these when framing the abstract, introduction, related work, hero figure, and before freezing the outline for a venue.
- NARRATIVE_REPORT.md or STORY.md with claims and evidence
- review-stage/AUTO_REVIEW.md (falls back to ./AUTO_REVIEW.md)
- Experiment results such as JSON files in figures/, screen logs or tables
- idea-stage/IDEA_REPORT.md if you used the idea-discovery pipeline (falls back to ./IDEA_REPORT.md)
- CLAIMS_FROM_RESULTS.md from the /result-to-claim step (preferred if available)
- The shared-references folder with writing-principles.md, venue-checklists.md and assurance-contract.md
Trigger the skill with natural requests
The skill's description lists trigger phrases. You can ask in English or Chinese. The source gives these examples: "写大纲", "paper outline", "plan the paper", "论文规划". You can also ask for a paper plan before you start writing.
The source also shows an optional override for the venue in the form /paper-plan "topic" — venue: NeurIPS. The default venue is ICLR. Supported venues are ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL and IEEE_CONF.
- "Plan the paper on my sparse attention results, targeting NeurIPS."
- "Write an outline for this paper from my review notes."
- "论文规划" or "写大纲" in the project folder
/paper-plan "topic" — venue: NeurIPSUnderstand the page budget and output
The skill sizes the outline to the venue's page limit. For ML conferences, the limit covers the main body through the conclusion, excluding references and appendix. ICLR and NeurIPS are set to 9 pages, ICML to 8, and AAAI to 7 technical-content pages plus references unless the current AAAI call says otherwise. For IEEE venues, references count toward the page total. The skill's section estimates are meant to sum to the maximum page count.
The result is an outline and planning notes. It is not finished paper text.
- Abstract: 150-250 words, with a self-contained check
- Introduction: contributions list and hero figure, about 1.5 pages
- Related Work: at least one page of synthesis, not a list
- Method and experiments: formal statements, figures and data sources
- Conclusion: restatement, limitations and future work
Safety notes: scripts and network
According to the skill's listing, paper-plan runs no scripts and uses no network. It works by reading and writing files in your project and reasoning over them.
The one external dependency is the review step. If you use the Codex-to-Claude review override, the reviewer role depends on a local claude-review MCP bridge. The skill's constants say REVIEWER_MODEL is claude-review, invoked through that bridge, and CLAUDE_REVIEW_MODEL can override the Claude model. Check that bridge's own setup before relying on it.
- Runs scripts: no
- Uses network: no
- Review override needs the local claude-review MCP bridge
Tips for better paper plans
The quality of the plan depends on the quality of your inputs. Run the claim-extraction step first if you have results, so the skill starts from adjudicated claims. The skill notes that if CLAIMS_FROM_RESULTS.md says verdict: REVIEW_UNAVAILABLE, the claims were not adjudicated; under assurance submission the skill stops and asks you to rerun /result-to-claim, while under assurance draft it continues and tags claims as unadjudicated.
Ask the agent to check the outline against the venue checklist before you freeze it, and keep your reviewer notes in the project so the plan can address known weaknesses.
- Put claims and reviewer notes in files, not only in chat
- Pick the venue before planning, since page budgets differ
- Treat the hero figure and claims-evidence table as your first drafting targets
What people use it for
- Planning a conference submission to a venue such as ICLR or NeurIPS from existing experiment results and reviewer notes
- Structuring an IEEE journal or conference paper, where references count toward the page limit
- Giving a student a section-by-section structure before drafting a first paper
- Keeping a paper plan aligned with earlier stages of the Auto-claude-code-research-in-sleep workflow
Frequently asked questions
Is the paper-plan skill free to use?
The skill is published under the MIT licence, as listed on this page. The MIT licence permits use under its terms. Check the repository's licence file for the full terms.
What do I need before I can use paper-plan?
You need project files with claims and evidence, such as NARRATIVE_REPORT.md or STORY.md, review-stage/AUTO_REVIEW.md, and experiment results. The skill also refers to a shared-references folder with three files. If you have none of these, the skill asks you to describe your contribution in 3-5 sentences, and the plan will be weaker.
Where do I install the paper-plan skill?
Unzip the downloaded skill into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings, Capabilities, Skills.
Does paper-plan write my full paper?
No. It produces an outline and planning notes, including section lengths, key citations, hero figure and a claims-evidence table. You still write the prose yourself.
Does paper-plan need the internet or run scripts?
The listing says it runs no scripts and uses no network. The only external piece is the optional Codex-to-Claude review override, which needs a local claude-review MCP bridge.
Which venues does paper-plan support?
The skill supports ICLR (the default), NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL and IEEE_CONF. You can name a venue in your request to override the default.
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 paper-plan if you already have experiment results and review notes and want a venue-aware outline before drafting. Skip it if you have no claims or results yet, or if you want the agent to write the full paper text, since it only produces an outline and planning notes.
Download the skill