A Packaged Skill for ML Paper Writing: Master-cai/Research-Paper-Writing-Skills

I keep two kinds of skills around for research work: ones that interrogate plans (grill-me, grill-with-docs — see the earlier post) and ones that interrogate writing. The Master-cai/Research-Paper-Writing-Skills repository is the cleanest entry I’ve found for the second category.

It packages Prof. Peng Sida’s open notes on paper writing into a single Skill that works across Claude Code, Codex, and Gemini. The point isn’t generic LaTeX advice — it’s a deliberate, ML/CV/NLP-flavored revision discipline that an agent can run on your draft.

What’s in the Box

The repository’s research-paper-writing/ directory has three layers:

research-paper-writing/
├── SKILL.md                 ← the workflow + global rules
├── references/              ← section-by-section playbooks
│   ├── abstract.md
│   ├── introduction.md
│   ├── related-work.md
│   ├── method.md
│   ├── experiments.md
│   ├── conclusion.md
│   ├── paper-review.md      ← adversarial self-review
│   ├── does-my-writing-flow-source.md
│   └── examples/
└── agents/openai.yaml

SKILL.md is the entry point — it tells the agent how to use the references. The references/ files are the actual writing playbooks the agent consults when it gets to a particular section.

Installation is just a copy:

Claude Code: ~/.claude/skills/ (global) or .claude/skills/ (project-local).
Codex: $CODEX_HOME/skills/.
Gemini: ~/.gemini/skills/.

After install you invoke it the way every Skill is invoked — by asking for it by name in the prompt.

The Five-Step Workflow

SKILL.md enforces a sequence that resists the temptation to start rewriting sentences first:

Clarify the narrative. Pin down problem → contributions → benefits → insights before touching any paragraph.
Apply section-specific guidance. Pull the appropriate playbook from references/ (e.g. introduction.md for the intro).
Rewrite paragraph-by-paragraph. One paragraph, one message, with a topic sentence that names the paragraph’s role.
Reverse-outline check. Read only the topic sentences in order — do they form a coherent spine? If not, the structure is wrong, not the prose.
Claim–evidence validation. Every claim, especially in the abstract and intro, is matched to a specific experimental result. Unsupported claims are either evidenced or weakened.

Two principles run through the whole thing:

“One paragraph, one message.” The clarity test asks three things: is there a single explicit message? does the opening sentence name the paragraph’s purpose? does each sentence connect via a logical relation (cause / contrast / consequence / refinement)?
Figures are core content. Visual quality is treated as part of the argument, not decoration that gets polished last.

The Introduction Playbook (representative example)

references/introduction.md is the file that shows what the whole skill is trying to do. Instead of vague “make the introduction compelling” advice, it gives a skeleton with substitutable templates per part:

The introduction has five parts:

Task and application context
Prior method limitations and root causes
Proposed solution and advantages
Additional contributions
Experimental validation

And then each part has named versions to choose from, e.g.:

Part	Templates
A. Task setup	(1) define niche task → applications · (2) lead with applications for a familiar task · (3) general task → specific setting · (4) expose technical challenge first via prior-method failures
B. Challenges	(1) chain existing methods → show progressive limitations · (2) classical insight → modern gap · (3) for novel tasks, decompose challenges into independent points
C. Method presentation	(1) single contribution, multiple advantages · (2) two sequential contributions for cascading challenges · (3) new module extending a prior pipeline · (4) observation-driven innovation

The most useful rule in this file is a warning: don’t hide method details behind abstract insights — papers that do this read as incremental even when the work is genuinely novel.

Each section playbook (method.md, experiments.md, related-work.md, …) follows the same pattern: a skeleton + a small menu of named templates + a checklist.

Adversarial Self-Review (`paper-review.md`)

The companion file paper-review.md flips the perspective: treat yourself as a hostile reviewer and probe every weak point before submission. It defines five rejection-risk categories:

Risk	What a reviewer attacks
Insufficient contribution	Problem is too common; the technique is already explored
Unclear writing	Missing technical details; modules introduced without motivation
Weak empirical effect	Marginal improvements; weak absolute performance
Incomplete evaluation	Missing ablations, baselines, challenging datasets
Problematic method design	Unrealistic settings; technical flaws; poor robustness; net negative value

Authors answer 25 specific questions across these five categories, marking each pass / needs revision / needs new experiment. The loop continues until no major rejection risk remains.

The hard constraint: every major claim, especially in Abstract and Introduction, must be both technically correct and explicitly supported by an experimental result. If unsupported, the claim is either backed by new evidence or rewritten weaker.

How It Composes With the Other Skills I Use

Stage	Skill	What it gates
Plan a piece of work	`grill-me`	Unstated assumptions in the plan
Plan inside a project’s domain	`grill-with-docs`	Undocumented decisions and glossary drift
Write the paper	`research-paper-writing`	Paragraphs without a message; claims without evidence

The three sit at different points in the same research pipeline: the first two enforce discipline before the work, the third enforces discipline after the work, while you’re turning results into prose.

Why I’m Adopting It

For paper writing I have always relied on ad-hoc heuristics (“does this paragraph have a topic sentence?”, “does the intro motivate the challenge?”). research-paper-writing turns those heuristics into:

a playbook an agent can apply consistently across sections,
a reverse-outlining check that makes structural problems visible before they become rewrite cycles,
a claim–evidence map that catches the single most common reason papers get desk-rejected — overclaiming in the abstract.

This pairs naturally with the research agent team rules I wrote down earlier: the team produces the LaTeX draft; this skill is the rubric the lead applies when the draft comes back.

References

Repository: Master-cai/Research-Paper-Writing-Skills
Original notes credited in the README: Prof. Peng Sida’s open study notes on academic writing.
License: MIT.