idea-battle
gyf9712/idea-battle
Claude Skill: three-track research idea advisor — stat methodology (Big Four), ML×stat (NeurIPS/ICML/ICLR/JMLR), medical AI (Nature family). Adapted from Punktheory/research-idea-and-battle.
Instalação
npx convez-ai gyf9712/idea-battle1 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
README
A tradução para o português deste README ainda está na fila.
Idea Battle — Three-Track Research Advisor (Claude Skill)
A Claude Skill that acts as your senior professor advisor, customized for a statistician working across three research tracks with different target venues. It identifies which track an idea belongs to, searches the current literature before speaking, stress-tests the idea against that track's actual gatekeepers, and — uniquely — evaluates the same idea's exit options across all three tracks.
Adapted from Punktheory/research-idea-and-battle (MIT).
English
The Three Tracks
| Track | Territory | Target venues |
|---|---|---|
| A — Stat Method | Statistical methodology & theory | JASA, Annals of Statistics, JRSS-B, Biometrika (the Big Four); Tier 2: EJS, Statistica Sinica, Biostatistics, JCGS |
| B — ML × Stat | Machine learning with statistical teeth (conformal, UQ, causal ML, learning theory) | NeurIPS, ICML, ICLR, COLT, AISTATS, JMLR; Tier 2: TMLR, UAI |
| C — Medical AI | Clinical AI & digital medicine | Nature Medicine, Nature Communications, npj Digital Medicine; staging: MICCAI, ML4H, MedIA |
Each track has its own reference file encoding the venue map, the contribution bar, what kills papers there (in the voice of the actual gatekeeper — a JASA AE, a NeurIPS Reviewer 2, a Nature Medicine editor), the search playbook, the graveyard patterns, and the clock speed.
What's different from the original skill
- Track-aware standards. The original speaks CV/top-conference dialect. This version knows that the Big Four, NeurIPS, and Nature Medicine judge the same idea by inverted standards (e.g., "rigorous first application" fails Track A but can be a Nature paper in Track C with strong enough validation).
- Positioning Across Tracks mode. One idea, three possible exits: a transfer-learning inference idea could become a Biometrika paper, a NeurIPS paper, or the methodological core of an npj Digital Medicine study. The skill assesses where it's strongest and what each repositioning demands.
- Track-specific graveyards. From assumption-laden high-dimensional inference (A) to UQ methods that never beat deep ensembles (B) to the COVID-prediction-model and Epic-sepsis graveyard (C).
- The statistician's edge, weaponized. In Tracks B and C, ideas are steered toward what a statistician can defend that a pure-CS group cannot: calibration, informative missingness, survival/competing risks, valid inference after selection.
Five core modes (inherited and re-calibrated)
Idea Evaluation · Battle Mode · Two-Tier Idea Generation · Landscape · Mentorship — plus Positioning Across Tracks.
Installation
# Global (available in all projects)
git clone https://github.com/gyf9712/idea-battle ~/.claude/skills/idea-battle
# Or project-level
git clone https://github.com/gyf9712/idea-battle /path/to/project/.claude/skills/idea-battle
Restart Claude Code — the skill triggers automatically on research-ideation conversations.
Usage
No special commands. Examples:
评估这个 idea:high-dimensional transfer learning 的统计推断,目标统计四大
Battle this idea: conformal prediction for LLM hallucination detection. Target NeurIPS.
我们有多中心 ICU 数据想做心衰 early warning,目标 npj Digital Medicine,帮我定位
这个 idea 投四大还是投顶会?
中文
三条赛道
为同时在三条赛道上工作的统计学研究者定制:
- 统计方法论 → 统计四大(JASA、AoS、JRSS-B、Biometrika)
- AI/ML × 统计 → ML 顶会(NeurIPS、ICML、ICLR、COLT、AISTATS)+ JMLR
- 医学 AI → Nature Medicine、Nature Communications、npj Digital Medicine
每条赛道有独立的 reference 文件:venue 画像、贡献标准、毙稿死因(以真实把关人的口吻——JASA AE、NeurIPS Reviewer 2、Nature Medicine 编辑)、文献搜索策略、墓地模式、时钟节奏。
与原版的区别
- 赛道感知的评判标准:同一个 idea 在三条赛道被倒置的标准评判(如"严格的首次应用"在四大不及格,但在 Nature 系可以是顶级论文)
- Positioning Across Tracks 模式:同一个 idea 评估三条赛道的出口,推荐最强的一条,并说明换赛道需要补什么
- 赛道专属墓地知识:从不可验证假设的高维推断(A),到从未打败 deep ensembles 的 UQ 方法(B),到 COVID 预测模型和 Epic sepsis 的集体翻车(C)
- 统计学家的比较优势:在 B、C 赛道把 idea 引向纯 CS 团队守不住的阵地(calibration、informative missingness、生存分析、selection 后有效推断)
安装
git clone https://github.com/gyf9712/idea-battle ~/.claude/skills/idea-battle
重启 Claude Code 即自动生效。聊到"评估 idea / battle / 投哪里 / 推荐方向"等话题时自动触发。
Structure
idea-battle/
├── SKILL.md # Core workflow: track detection → search → snapshot → lenses → modes
├── references/
│ ├── stat-method.md # Track A: Big Four profile
│ ├── ml-stat.md # Track B: ML venues + JMLR profile
│ └── medical-ai.md # Track C: Nature-family profile
└── evals/evals.json # Test prompts used during development
License
MIT. Derivative of Punktheory/research-idea-and-battle (MIT).