Paper digger ideate
Skill Iu1ky/paper-digger/plugins/paper-digger/skills/paper-digger-ideate
Use when generating and selecting research ideas from constraints, literature gaps, and novelty/feasibility/impact scoring. Triggers on 选题, 找研究方向, idea发掘, research ideation.From its SKILL.md
npx -y skills add Iu1ky/paper-digger --skill paper-digger-ideateAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 16 days oldThe repository was created 16 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.6 KB, 905 tokens by cl100k_base, as published. Nobody here has run it
Paper Digger Ideate — 方向确定与 idea 发掘(Phase 1)
模糊前端:从研究者画像出发,发散生成候选 idea,收敛到 1–2 个值得做的。确定前必须跑评价节点①(paper-digger-evaluate),再由人在 checkpoint 拍板。
运行时与检索
定位同级 paper-digger skill,把它的 scripts/ 加入 Python import path 后调用 paper_digger.*。使用宿主可用的网页/学术检索工具并优先原始论文、官方索引和 DOI;没有网页能力时,不做新颖性确认,只输出待检索问题和 blocker。
输入
研究者画像(领域、兴趣、约束:算力?湿实验?时间?目标档次),来自 00_profile/ 或现问。
发散(独立多角度)
按 state.effort 选择预算;不要仅因宿主支持 subagent 就全量并行:
lean:2 个差异最大的 lens,最多 4 张卡。standard:3 个 lens,最多 6 张卡。deep:仅在用户明确要求、top 候选接近或新颖性未决时使用全部 5 个 lens。
可选 lens:
- gap-driven(文献空白)· method-transfer(他领域方法迁移)· cross-disciplinary(交叉)· contrarian(反主流假设)· trend-driven(新兴趋势)
用宿主网页/学术检索能力落地;有学术 API(见 state.capabilities)时增强。每个候选产出一张 idea card:
{id, one_liner, novelty_claim, why_now, required_evidence, feasibility_note, risk, scores: {novelty, feasibility, impact, fit}}(各 0–5)。
收敛
- 打分排序:
paper_digger.ideate.save_idea_cards(workspace, cards)—— 内部rank_ideas按 NFIF 加权(NFIF_WEIGHTS),写01_ideation/idea_cards.md。 - 用一次 Devil's-advocate 复核 top 候选:真的新吗?是否已被做过?是否非真问题?
- 评价节点①:对 top 1–2 候选跑
paper-digger-evaluate(node=1,重 A1 价值 / A2 新颖 / B4 思维固定),拿 verdict。 - checkpoint:呈现排序 + verdict + 必修项,由人选定 1(或 2)个。
- 落定:
paper_digger.ideate.confirm_idea(workspace, idea_id, summary=...)—— 写01_ideation/confirmed_idea.md并置state.decisions.idea。
铁律
- 不臆造:新颖性主张要有文献依据;拿不准就标注、去查,绝不编。
- 节点①必过:确定 idea 前必须跑评价;RED 强烈建议换框架/换题(人拍板)。
- 发散阶段保持多样性,避免一上来锚定单一框架(对应 B4 思维固定)。
What ships with it: 1 file
231 B alongside SKILL.md
agents/
- openai.yaml231 B