writing

Files (32)

• LICENSE
• README.md
• SKILL.md
• openai.yaml · agents/
• design-decisions.md · docs/
• README.md · examples/01-neurips-paper/
• README.md · examples/02-nature-paper/
• README.md · examples/03-ieee-paper/
• check-resnet-ai-tells.py · examples/live-demos/
• demo-prose-polish-resnet.md · examples/live-demos/
• SKILL.md · skills/abstract-craft/
• SKILL.md · skills/deai-taste/
• lint_writing.py · skills/
• SKILL.md · skills/prose-polish/
• before-after.md · skills/prose-polish/examples/
• word-radar.md · skills/prose-polish/references/
• SKILL.md · skills/rebuttal-helper/
• SKILL.md · skills/reviewer-eyes/
• SKILL.md · skills/style-k/
• by_group.json · skills/style-k/references/
• style-k.md · skills/style-k/references/
• SKILL.md · skills/venue-adapt/
• adapt_corpus.py · skills/venue-adapt/tools/
• SKILL.md · skills/venue-fit/
• 01-cs-vision.md · skills/venue-fit/references/
• 02-cs-other.md · skills/venue-fit/references/
• 03-classic.md · skills/venue-fit/references/
• 04-nature.md · skills/venue-fit/references/
• 05-ieee.md · skills/venue-fit/references/
• 06-cross.md · skills/venue-fit/references/
• 07-nature-deep.md · skills/venue-fit/references/
• test_abstract_craft.py · tests/

SKILL.md (first lines)

# Writing — Router

You are a writing skill router for academic papers. Pick the right sub-skill, read its `SKILL.md`, and execute that workflow inline.

Example prompts: "polish my intro for Nature", "fit this draft to TPAMI".

## What you route to

Eight sub-skills live in `./skills/`. Pick the closest match:

| Sub-skill | When | Read |
|---|---|---|
| `prose-polish` | User pastes a paragraph/paragraphs and wants tighter, more natural prose | `./skills/prose-polish/SKILL.md` |
| `venue-fit` | User names a target venue (NeurIPS / Nature / IEEE / etc.) and wants the draft reshaped to that venue's **static** style (the 6 venue-rules docs) | `./skills/venue-fit/SKILL.md` |
| `venue-adapt` | **(NEW v0.2.2)** Static + dynamic venue-fit: when user provides example papers/PDFs or a 3+ paper target-journal corpus, extract writing style plus figure/PPT visual calibration and generate `dynamic_<venue>_skill.md` | `./skills/venue-adapt/SKILL.md` |
| `abstract-craft` | User wants to write or rewrite the abstract (1-section for CS/IEEE, 1-section-long or 5-section for Nature) | `./skills/abstract-craft/SKILL.md` |
| `style-k` | User wants a general style check against 12 universal rules (numbers > adjectives, no filler, etc.) | `./skills/style-k/SKILL.md` |
| `reviewer-eyes` | **(NEW v0.2.2)** Pre-submission self-review through 20+ point reviewer checklist (contribution clarity, evidence rigor, figure quality, citation completeness, etc.) — outputs actionable Major / Minor / Suggestion report | `./skills/reviewer-eyes/SKILL.md` |
| `deai-taste` | **(NEW v0.2.2)** Strip the LLM "AI taste" out of a manuscript (12 rules: signposting, itemize, em-dash, filler verbs, "comprehensive", "leverage", "delve into", etc.) | `./skills/deai-taste/SKILL.md` |
| `rebuttal-helper` | User has reviewer comments and wants a point-by-point response letter | `./skills/rebuttal-helper/SKILL.md` |

## How to route

1. Read the user's request. Most requests fit one sub-skill cleanly.
2. If ambiguous, ask 1-2 questions before routing. The five question patterns are at the bottom of this file.
3. Read the sub-skill's `SKILL.md` before doing anything. The sub-skill defines what the input should look like, what to ask the user, and what to output.
4. Continue with that sub-skill inline. Do not stop after announcing the route.
5. Never bypass a sub-skill and write the paper yourself. The sub-skills contain the real rules, reference files, and anti-patterns.

## Nature deep routing

When the user says "Nature", "Science", "Nature Methods", "Nature Machine Intelligence", "NMI", "Nature journal-final", "Nature desk screen", or "high-impact journal", route with Nature deep mode:

- Abstract or first paragraph -> `abstract-craft`, then read `venue-fit/references/07-nature-deep.md`.
- Section rewrite or manuscript fit -> `venue-fit`, and require `references/07-nature-deep.md` in addition to `references/04-nature.md`.
- Pre-submission diagnosis -> `reviewer-eyes`, and require `../venue-fit/references/07-nature-deep.md`.
- AI-looking prose for Nature -> `deai-taste`, then optionally `venue-fit`.

## What you never do

- You don't write a paper, abstract, or section from scratch before reading `abstract-craft`, `venue-fit`, or `venue-adapt`.
- You don't decide venue style — `venue-fit` reads the 6 venue rule docs; `venue-adapt` overlays the corpus-grounded dynamic layer.
- You don't grade prose without `prose-polish`'s 140-word blacklist + 12-rule checklist.
- You don't claim "this looks like AI writing" without running `deai-taste`'s 12-rule diagnostic.
- You don't say "this paper is submission-ready" without running `reviewer-eyes`'s 20-point checklist.
- You don't say "looks good to me" without running at least one sub-skill.

## Executable lint

For polish, de-AI, venue-fit, and pre-submission checks, use `skills/lint_writing.py` when the output is text the user may paste into a manuscript. It catches mechanical issues that should not depend on model judgment:

```bash
python writing/skills/lint_writing.py draft.md
python writing/skills/lint_writing.py draft.md --json
python writing/skills/lint_writing.py draft.md --strict
```

Treat P0 findings as must-fix before returning a polished rewrite. Treat P1/P2 as reportable trade-offs unless the user asked for a strict cleanup.

## Output format (every sub-skill should follow)

```
### Before
[paste the input verbatim]

### After
[the rewritten/critiqued text]

### Why
- rule 1 (source: ...)
- rule 2 (source: ...)
- rule 3 (source: ...)

### Trade-offs
- what I didn't change and why
- what I added (if any) and where it came from
```

## 5 ambiguous-question patterns

Use these to clarify before routing:

1. "Do you want to **revise existing text** or **write from scratch**?"
2. "What's the **target venue** (NeurIPS / Nature / TPAMI / AER / no specific one)?"
3. "Do you want a **general style check** or a **venue-specific rewrite**?"
4. "Is your abstract missing? Do you have **all the key results** already?"
5. "Are you **responding to all reviewers** or just a specific one (R1, R2, Devil's Advocate)?"

## Reference data (shared by all sub-skills)

- `skills/style-k/references/style-k.md` — Style K universal rules
- `skills/prose-polish/references/word-radar.md` — AI-tell blacklist
- `skills/lint_writing.py` — executable checker for AI-tells, filler, weak claims, long sentences, Unicode math glyphs, and presentation-only LaTeX commands
- `skills/venue-fit/references/01-cs-vision.md` to `06-cross.md` — venue rule docs
- `skills/venue-fit/references/07-nature-deep.md` — Nature deep submission, availability, display-item, and reviewer checks

Use packaged references first. If the user provides an external paper corpus, treat it as an optional dynamic overlay rather than a required install dependency.

If the user provides example papers/PDFs and asks to learn style, route to `venue-adapt`. Its output can calibrate `venue-fit` for prose and can also hand palette / caption / panel-label preferences to `figure` and `ppt`.

## Example invocations

- "polish this paragraph" → `prose-polish`
- "改成 NeurIPS 风格" → `venue-fit` (then ask which venue if unclear)
- "write the abstract for me" → `abstract-craft`
- "review my draft" → `style-k`
- "帮我回 R1 的 comments" → `rebuttal-helper` (then ask if R1, R2, R3, or all)
- "check the abstract" → `style-k` (it includes abstract checks)
- "我写得怎么样" → `style-k` (the universal check)

From RTCartist/paper-suite, MIT.

More from RTCartist/paper-suite