Embodied AI Paper Writer
OpenGHz/embodied-ai-paper-writer
Coach corpus-tuned para papers de robótica em qualquer agente LLM
Ver no GitHubSkill portável baseado em corpus de 63 papers de top-tier (CoRL, RSS, ICRA, IROS, Science Robotics 2022–2026) que ensina o ofício de escrever papers em robótica e IA embodied. Funciona com qualquer agente LLM capaz de carregar contexto markdown, fornecendo padrões quantificados para título, abstract, método, experimentos e conclusões.
Instalação
npx convez-ai OpenGHz/embodied-ai-paper-writer1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Estruturar abstracts e introduções de papers de robótica
- Organizar seções de método, trabalhos relacionados e experimentos
- Padronizar captions de figuras, tabelas e convenções de escrita
- Treinar agentes LLM como coaches de escrita científica em IA embodied
Veredito

Skill altamente especializada e bem fundamentada em dados reais (63 papers analisados). Excelente para pesquisadores de robótica que usam agentes LLM, mas requer conhecimento prévio do domínio. Diferencia-se por regras quantificadas em vez de conselhos genéricos.
README
A tradução para o português deste README ainda está na fila.
Embodied AI Paper Writer
English | 简体中文
A portable agent skill (SKILL.md + reference playbooks) that coaches the writing craft of embodied-AI papers — distilled from 63 top-conference papers across CoRL, RSS, ICRA, IROS, and Science Robotics (2022–2026). Works with any LLM agent that can load markdown context.
What this is
A drop-in agent skill that turns any modern LLM agent — Claude Code, Cursor, Continue, Cline, Aider, the Anthropic / OpenAI SDKs, or anything else that can load a system prompt — into a corpus-tuned coach for writing robotics / embodied-AI papers. It teaches:
- Title patterns, abstract moves, intro arcs.
- Method / Related Work organization.
- Experiments setup framing, results paragraph rhythm, ablation narration.
- Figure roles, caption templates, table conventions.
- Conclusion / Limitations / Future Work / Appendix.
- Section openers, pivots, connectors, the contribution-restatement spiral.
- A phrasebook of openers, hedges, and anti-patterns.
It does not decide which experiments to run, validate technical claims, suggest research directions, translate, or run your LaTeX build. See SKILL.md for the full boundary list.
Why a skill, not a textbook
Most writing advice is generic ("be clear, cite related work"). This skill is quantified and corpus-tuned:
- "Abstract = 120–250 words, 5 moves (Frame → Gap → Contribution → Method → Results)."
- "Method tense = present, system-as-subject."
- "Plot caption = mean ± StdErr + sample size, always."
- "Contribution noun phrase repeats 5–7× verbatim across the paper."
- "F1 teaser caption = 3–6 sentences; F3 hardware caption = 1 sentence."
Each rule traces back to observed patterns in the 63-paper corpus.
Quick start
1. Install
Copy SKILL.md and references/ into wherever your agent loads system prompts or skill files. Examples:
- Claude Code — user-level:
~/.claude/skills/embodied-ai-paper-writer/ - Claude Code — project-level:
<your-project>/.claude/skills/embodied-ai-paper-writer/ - Cursor / Continue / Cline / Aider — paste
SKILL.mdinto the custom-rules / system-prompt panel; keepreferences/next to it and attach files on demand per the routing table. - Custom agent (Anthropic / OpenAI / local model SDKs) — load
SKILL.mdas the system message; lazily readreferences/*.mdbased on the routing table. - Plain chat — paste
SKILL.mdinto the conversation; follow up with the relevant playbook when the routing table calls for it.
# Example: install as a user-level skill for Claude Code
SKILL_DIR="$HOME/.claude/skills/embodied-ai-paper-writer"
mkdir -p "$SKILL_DIR"
cp -r SKILL.md references "$SKILL_DIR/"
The skill is just markdown + frontmatter — nothing in the runtime behavior is locked to one vendor.
2. Invoke it
Once loaded, the agent will engage on phrases like:
- "Help me write the abstract for my CoRL submission."
- "Caption this figure — it's a 3-panel success-rate plot."
- "My Limitations section sounds defensive — fix it."
- "Review my paper's arc."
- 「帮我润色一下这段 Intro」
- 「rebuttal 怎么写」
See the full trigger list in SKILL.md (frontmatter description).
Project layout
embodied-ai-paper-writer/
├── SKILL.md # Operating manual (entry point)
├── references/
│ ├── titles.md # Title patterns and architectures
│ ├── abstract-intro-playbook.md # Abstract + Intro construction
│ ├── method-relatedwork-playbook.md # Method + Related Work
│ ├── experiments-results-playbook.md # Experiments, Results, Ablations
│ ├── figures-tables-playbook.md # F1–F8 figure roles, caption templates
│ ├── language-phrasebank.md # Rhetorical phrasebook (A–K)
│ ├── flow-transitions.md # 6-move paper arc, pivots, connectors
│ ├── closing-appendix-playbook.md # Conclusion / Limitations / Appendix
│ └── research/ # Raw extraction (read for traceability)
│ ├── _paper_roster.md # 63-paper corpus index
│ └── 00–08 + topical extracts
├── LICENSE
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── SECURITY.md
├── SUPPORT.md
├── CHANGELOG.md
└── CITATION.cff
The skill has two layers:
| Layer | Files | Purpose |
|---|---|---|
| Operational | SKILL.md + 8 playbooks in references/*.md | Loaded by the agent on demand via the routing table in SKILL.md |
| Research | 9 files in references/research/ | Raw evidence — read only when traceability is needed |
The operational playbooks are 8–25 KB each. The research files (50–200 KB) exist so any rule can be traced back to source patterns; they are not loaded during normal use.
How the skill thinks
The agent follows five execution scenarios (defined in SKILL.md):
| Scenario | Trigger | Behavior |
|---|---|---|
| A — Write new section | "draft my intro" | Confirm scope → gather briefing → load reference → pre-draft checkpoint → draft → self-review → deliver with 3-line summary |
| B — Fix/review existing | "fix my method" | 4-layer diagnostic scan → numbered findings → checkpoint before rewrite |
| C — Question only | "how long should X be?" | Answer concisely with the number + the rule; don't volunteer to rewrite |
| D — Caption a figure | "caption this plot" | Identify F1–F8 type → confirm role → template → verify panel notation + task-name consistency + length budget |
| E — Whole-paper arc | "review my paper" | 4-lens check (arc / spiral / tense / figure coupling) → prioritized fix list |
Plus 13 universal rules including: lock the contribution noun phrase, disclose deltas not absolutes, one pivot per gap, every limitation pairs with future-work mitigation, and pushback escalation policy.
Honest boundaries
- Sample: 63 papers, 2022–2026, CoRL / RSS / ICRA / IROS / Science Robotics. NeurIPS / ICML robotics tracks are under-represented. CVPR-adjacent robotics is under-represented.
- Anglophone bias: corpus is English-language.
- Conventions evolve: re-extract after mid-2027.
- Writing-craft only: this skill cannot judge whether a paper is publishable, whether the contribution is strong, or whether the experimental design is sound. It can only judge whether the writing follows the conventions of papers that did get published.
- Defaults calibrated to CoRL: when writing for ICRA / IROS / Science Robotics, the skill asks once to recalibrate venue-specific conventions.
Contributing
Pattern corrections, new corpus additions, and anti-pattern submissions are welcome. See CONTRIBUTING.md.
Common contributions:
- Add a paper to the corpus → see the Paper Addition template.
- Report a wrong rule → see the Bug Report template.
- Suggest a missing pattern → see the Pattern Suggestion template.
Citation
If this skill helped your writing, please cite the repository. See CITATION.cff.
License
MIT — use, fork, adapt, redistribute. Attribution appreciated, not required.
Acknowledgments
This skill was distilled using the Nuwa · Skill造人术 methodology — a structured pipeline for turning expert corpora into operational skills (SKILL.md + reference layer). The methodology made the difference between "generic writing advice" and "corpus-tuned quantified rules."
Original corpus papers and authors retain all rights to their work; only writing patterns and conventions are abstracted here.