How to use the Paper Outline Planner agent skill (paper-plan)

What an agent skill is and what paper-plan makes the agent do

An agent skill is a folder containing a SKILL.md file with instructions the agent loads when a request matches the skill's description. This one is named paper-plan and is triggered when you ask for an outline, a paper plan or the planning of a paper before writing.

When triggered, the agent reads your existing research notes, review conclusions and experiment results, then drafts a structured outline of the paper section by section. It respects page limits and formatting norms for a chosen venue such as ICLR, NeurIPS or IEEE journals. It can optionally use a reference paper to copy the structural layout (section order, subsection density, figure count) without copying its wording or claims.

The skill produces an outline and plan, not finished paper text. The outline still needs your judgement on the claims.

• Inputs it looks for: NARRATIVE_REPORT.md or STORY.md, review-stage/AUTO_REVIEW.md (falls back to ./AUTO_REVIEW.md), experiment results such as JSON files in figures/, and idea-stage/IDEA_REPORT.md if present.
• Default venue is ICLR. You can override it, for example by adding venue: NeurIPS to your request. Supported venues include ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM, IEEE_JOURNAL and IEEE_CONF.
• If none of the input files exist, the skill asks you to describe the paper's contribution in 3-5 sentences.

Install the paper-plan skill from the ZIP

Download the ZIP for skills/paper-plan from this page and unzip it. Place the resulting paper-plan folder, which contains SKILL.md, in one of two locations: ~/.claude/skills/ to make it available across your projects, or a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/paper-plan/
• Project install: <your-project>/.claude/skills/paper-plan/
• Claude apps: Settings → Capabilities → Skills, then upload the skill.

What the skill needs before it runs

The skill needs an environment where the agent can run Bash, read and write files, and use web tools. It is written for Claude Code or a SKILL.md-compatible agent. Your working directory should contain project files such as NARRATIVE_REPORT.md, AUTO_REVIEW.md or experiment JSON results, because the output depends on them.

Some pieces are optional. Cross-model review uses Codex MCP or other reviewer MCP servers such as mcp__codex or mcp__claude-review. Python 3 and the extract_paper_style.py helper are needed only if you use the --style-ref option.

• Required: Claude Code or a SKILL.md-compatible agent with Bash, file read/write and web tools enabled.
• Optional: Codex MCP or another reviewer MCP server for the outline review step.
• Optional: Python 3 and extract_paper_style.py, only for the style-reference option.
• Recommended: project files such as NARRATIVE_REPORT.md, AUTO_REVIEW.md or experiment JSON results in the working directory.

Trigger the skill with a natural request

You do not need a special command. Ask in plain language, and include the topic or the narrative document if it is not obvious from the working directory. The skill's description lists these trigger phrases: 写大纲, paper outline, plan the paper and 论文规划.

The skill also accepts an argument hint of the form [topic-or-narrative-doc] followed by an optional style reference. Venue can be set in the request as shown in the Constants section of the skill.

• "Plan the paper for my results. Target NeurIPS."
• "Write an outline for this project using NARRATIVE_REPORT.md."
• "写大纲" or "论文规划" when you want the plan before writing.
• "Make a paper plan against the ICLR 9-page limit."

Use the optional style reference

Add a style reference to steer the structure of the outline toward an existing paper. It is off by default. The reference can be a local TeX directory or file, a local PDF, an arXiv id, or an http(s) URL. Overleaf URLs and project IDs are rejected; clone the project first and pass the local path.

The skill uses the reference only for structural guidance such as section order, subsection counts, theorem density and figure budget. It is instructed not to copy prose, claims, examples, section names or terminology from it. When a style reference is used and your project has figures or results, the skill also writes a GAP_REPORT.md that marks each expected section slot as covered, partial or missing.

• Input forms: local TeX dir or file, local PDF, arXiv id (2501.12345 or arxiv:2501.12345), http(s) URL.
• Output: GAP_REPORT.md, with each slot classed as covered, partial or missing.

Review the outline and check it against the venue

The skill may invoke an external reviewer model to critique the outline. The source shows this as an optional step, and the file does not show the whole workflow, so do not assume a reviewer pass always happens. Treat the outline as a draft.

Check that section lengths fit the venue's page limit. For IEEE venues, references count toward the page limit. Check every claim in the outline against your own results, because the skill cannot verify them for you.

• ICLR and NeurIPS: main body is limited to 9 pages in the skill's defaults, excluding references and appendix.
• ICML: 8 pages. AAAI: 7 technical-content pages plus references unless the current call for papers says otherwise.
• IEEE journal Transactions: about 12-14 pages total. Letters: about 4-5 pages. IEEE conference: about 5-8 pages, including references.

Safety notes: scripts and network access

This skill runs scripts and uses the network. Its tool list includes Bash, WebSearch and WebFetch, and it can call reviewer MCP servers. The style-reference helper is a Python script that fetches the reference source, so only pass sources you trust.

Review the SKILL.md file before installing, and be aware that reviewer calls send your outline to the external model you configure. Do not pass unpublished results to a reviewer you have not approved for that purpose.

• Runs scripts: yes. Uses network: yes.
• The style-reference option resolves sources from local paths, arXiv ids or URLs.
• Reviewer MCP servers are optional and depend on external model access, which may be unavailable.