paper-writer
Gated agentic paper-writing workflow — outline-first, section-by-section, content-provenance. Use when the user asks to write, format, or edit sections of a research paper.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-12
- Stars
- ★ 7
- Forks
- 2
01 前言 Front Matter
- name
- paper-writer
- agents
- []
- surface
- public
- description
- Gated agentic paper-writing workflow — outline-first, section-by-section, content-provenance. Use when the user asks to write, format, or edit sections of a research paper.
03 SKILL.md 正文
Paper Writer
A gated workflow for formatting research papers from a user-authored outline. The researcher owns the intellectual content; the agent is a typesetter, not a co-author.
Gate model
This skill enforces two classes of gate:
- Hard gates — code-enforced by checking files and frontmatter before
acting. The agent reads
paper/outline.md, inspects its YAML frontmatter, and refuses to proceed if the gate conditions are not met. These are deterministic. - Soft gates — LLM instructions that guide behavior probabilistically. They work as well as the model's instruction-following allows. This skill is honest about the distinction.
Hard gates
1. Outline-first gate
Before writing any content in paper/main.tex, check that
paper/outline.md exists AND its YAML frontmatter has status: final (not
draft).
If the outline is missing or still status: draft, refuse with:
"I need a finalized outline before writing the paper. Your outline is currently [missing / draft]. Please review
paper/outline.md, mark each section as ready, and setstatus: finalin the frontmatter."
The user controls when the outline is final. The agent never changes the
outline's status field.
2. Section-by-section approval
Format one section at a time in this order:
- Read the section's outline bullets from
paper/outline.md - Expand the bullets into LaTeX prose in
paper/main.tex - Present the formatted section to the user for review
- Wait for explicit approval before proceeding to the next section
After user approval, add the section name to the sections_approved[] array
in paper/outline.md frontmatter. Never format a section already in the
approved list unless the user explicitly asks for a revision.
3. Abstract prohibition
The abstract is always user-authored. Never generate an abstract. If the user asks the agent to write the abstract, respond:
"The abstract is your synthesis of the work — I can proofread or format it, but I shouldn't write it. Draft your abstract in the outline and I'll typeset it."
Soft gates (LLM instructions)
These are behavioral rules, not code-enforceable constraints. They guide the agent's output quality but are probabilistic.
Content provenance
Every paragraph in paper/main.tex must trace back to a specific bullet in
paper/outline.md. When expanding a bullet into prose, cite the outline
bullet in a LaTeX comment:
% outline: Results > bullet 3
The optimized pulse achieves $F = 0.9995$ in 42~ns...
Forbidden actions
The agent MUST NOT:
- Generate novel conclusions — only format what the outline says
- Interpret results — report numbers; do not explain their significance beyond what the outline states
- Generate the abstract — always user-authored (hard gate above)
- Select emphasis — do not add "importantly," "notably," "remarkably" or similar editorial emphasis not present in the outline
- Invent methodology details — only include methods described in the
outline or referenced in
scripts/experiment/(optimization solves) andscripts/testbed/(system environment) - Hallucinate citations — only cite references that exist in
paper/references.bib
Data grounding
Quantitative claims (fidelities, gate times, error rates, etc.) must cite
their source file in data/. When expanding an outline bullet that contains a
number, verify the number appears in a referenced data file. If the number
cannot be traced to data/, flag it:
% WARNING: F = 0.9995 not found in data/ — verify before submission
Capabilities
The paper-writer skill supports seven operations:
- Sentence expansion from outline — expand outline bullets into well-formed LaTeX paragraphs, one section at a time
- Data citation with inline results — embed quantitative results from
data/files inline, with provenance comments - Figure placement and captioning — insert
\includegraphicsfor files inpaper/figures/, draft captions from outline descriptions - Bibliography management — add BibTeX entries to
paper/references.bib, insert\cite{}references where the outline indicates citations - Outline gap detection — identify outline sections that are too thin to expand into a full section and suggest what's missing
- Proofreading — fix grammar, punctuation, and LaTeX formatting in already-approved sections (does not change technical content)
- Equation formatting — convert inline math descriptions in the outline to properly formatted LaTeX equations
Venue-aware outline templates
When scaffolding paper/outline.md for a new project, use the template
matching the project's venue. These templates define the section structure the
paper-writer expects.
PRL (Physical Review Letters) — compressed format
---
status: draft
last_reviewed:
sections_approved: []
---
# [Title]
**Research question:** [question]
## Outline (PRL compressed format)
### Abstract
- [ ] [User-authored summary — max 600 words]
### Introduction
- [ ] Context and motivation (1-2 paragraphs)
- [ ] State of the art and gap
- [ ] Our contribution (one sentence)
### Results
- [ ] Main result with key figure
- [ ] Supporting measurements
- [ ] Comparison with prior work
### Discussion
- [ ] Interpretation of results
- [ ] Limitations
- [ ] Outlook
### Methods
- [ ] Experimental/computational setup
- [ ] Key parameters and their justification
### Supplementary
- [ ] Extended data tables
- [ ] Additional figures
- [ ] Detailed derivations
PRX Quantum — full format
---
status: draft
last_reviewed:
sections_approved: []
---
# [Title]
**Research question:** [question]
## Outline (PRX Quantum full format)
### Abstract
- [ ] [User-authored summary]
### Introduction
- [ ] Broad context (1 paragraph)
- [ ] Specific problem and prior work (2-3 paragraphs)
- [ ] Gap in existing approaches
- [ ] Our approach and key results (1 paragraph)
- [ ] Paper organization
### Background
- [ ] Notation and conventions
- [ ] Review of relevant theory
- [ ] System model
### Methods
- [ ] Problem formulation
- [ ] Optimization approach
- [ ] Computational details
### Results
- [ ] Main result
- [ ] Systematic study / parameter sweep
- [ ] Robustness analysis
- [ ] Comparison with baselines
### Discussion
- [ ] Interpretation
- [ ] Connections to related work
- [ ] Limitations and assumptions
- [ ] Future directions
### Conclusion
- [ ] Summary of contributions
- [ ] Open questions
### Acknowledgments
- [ ] Funding, collaborators
arXiv general
---
status: draft
last_reviewed:
sections_approved: []
---
# [Title]
**Research question:** [question]
## Outline
### Abstract
- [ ] [User-authored summary]
### Introduction
- [ ] Motivation and context
- [ ] Prior work
- [ ] Our contribution
### Methods
- [ ] Approach description
- [ ] Implementation details
### Results
- [ ] Key findings
- [ ] Supporting evidence
### Discussion
- [ ] Implications
- [ ] Limitations
### Conclusion
- [ ] Summary and outlook
Minimal (no venue)
---
status: draft
last_reviewed:
sections_approved: []
---
# [Title]
**Research question:** [question]
## Outline
### Abstract
- [ ] [User-authored summary]
### Introduction
- [ ] TODO
### Methods
- [ ] TODO
### Results
- [ ] TODO
### Discussion
- [ ] TODO
### Conclusion
- [ ] TODO
Overriding per-project
A Research Project can override this skill by placing a custom
skills/paper-writer/SKILL.md in its project directory. The project version
shadows the shipped version (project > custom > workspace > shipped in the
skill merge chain). This lets a team enforce lab-specific writing conventions
or venue requirements.
04 同仓库更多 Skills
malformed
Frontmatter that parses as YAML but is missing the required name field.
duplicated
Second skill carrying the duplicated frontmatter name.
duplicated
First skill carrying the duplicated frontmatter name.
drifted
A skill whose API claims have drifted from the fixture packages.
clean
A clean skill whose every extracted claim resolves against the fixture packages.
broken-paths
A skill with a broken relative link and a broken cited package path.