ai-paper-pipeline（AI 纸管道）

AI Paper Pipeline

Turn a rough paper idea or a long "mega prompt" into a reusable, reality-grounded paper project scaffold.

What this skill should do

• Normalize a user's long paper-workflow prompt into a maintainable skill/project structure.
• Keep the main workflow concise in SKILL.md and push bulky reference text into references/.
• Preserve academic-integrity constraints: no fabricated experiments, no fake citations, no unsupported claims.
• Prefer creating reusable project scaffolding over dumping one giant prompt blob.

Default workflow

1. Identify whether the user wants one of these:

- skill cleanup / packaging for the paper workflow itself - project initialization for a specific paper - template ingestion from a pasted mega prompt

1. If the user pasted a large workflow prompt, extract and organize it into:

- SKILL.md for concise usage instructions - references/ for long-form reference content - templates/ for starter files like RESTRICTS.example.yaml

1. Keep only trigger logic, workflow guidance, and file navigation in SKILL.md.
2. Put long source material, detailed prompts, and heavy policy text in references/.
3. If the user wants a paper project initialized, create at minimum:

- MEGA_PROMPT.md - RESTRICTS.yaml - PROGRESS.md - plans/ - code/, data/, docs/, results/ - paper/mypaper/main.tex - paper/mypaper/sections/

1. After edits, package or commit changes if appropriate.

File layout for this skill

ai-paper-pipeline/
├── SKILL.md
├── MEGA_PROMPT.md
├── references/
│   ├── full-pipeline-template.md
│   └── project-scaffold.md
└── templates/
    └── RESTRICTS.example.yaml

When to read extra files

• Read MEGA_PROMPT.md when you need the concise built-in version of the 25-stage workflow.
• Read references/full-pipeline-template.md when the user wants the verbose original template or asks to reconstruct/port the full prompt.
• Read references/project-scaffold.md when the user wants to initialize a concrete paper project directory.
• Read templates/RESTRICTS.example.yaml when initializing a new paper project or drafting a restrictions file.

Working rules

• Treat the paper as a real research artifact, not a vibe-writing exercise.
• Never claim experiments, datasets, ablations, or statistical tests that are not actually present.
• Never keep huge duplicated prompt text in multiple files.
• Prefer editable project artifacts over giant single-message outputs.
• Keep the paper workflow cyclical: literature → design → run → analyze → draft → review → revise.

Good outputs

A. User says: "整理成一个 Skill"

Do this:

• Clean up the current skill folder.
• Convert ad-hoc text into proper SKILL.md + references/ + templates/.
• Keep SKILL.md concise and reusable.

B. User says: "按这个模板起一个论文项目"

Do this:

• Create a new <project>-paper/ scaffold.
• Copy in starter files.
• Replace placeholders with project-specific metadata where provided.

C. User says: "把这份 mega prompt 落库"

Do this:

• Save the raw template in references/ or project root.
• Avoid bloating SKILL.md with the full raw text.

Final step

After modifying this skill or creating paper-project files in the workspace, commit the changes with a clear git message.