| name | idea-research-skill |
| description | Research idea evaluation and iteration assistant covering the six-stage LLM Wiki pipeline — paper discovery (arXiv/Semantic Scholar), ingest to raw/, Claude Code compiles raw/ to a structured wiki/, Obsidian visualization, multi-model debate (Claude/Gemini/Codex CLI), and Perplexity novelty validation. Use when the user wants to register a new idea, compile a paper wiki, run a multi-model debate, or validate an idea's novelty. Triggered by keywords like "idea validate", "paper wiki", "compile raw", "novelty check", "多模型讨论", "idea 验证", or direct invocation. |
Idea Research Skill
Six-stage LLM-Wiki pipeline for research idea evaluation.
Workspace Layout
Everything lives under the repo-root idea-research/ directory (gitignored):
idea-research/
├── ideas/ one subdir per idea, slug-named
│ └── <idea-slug>/
│ ├── idea.md statement, hypothesis, evaluation criteria
│ ├── raw/ source papers in markdown (append-only)
│ ├── wiki/ compiled concept/paper notes (Claude output)
│ │ ├── papers/ paper-level summaries
│ │ ├── concepts/ cross-paper syntheses
│ │ └── gaps/ identified open problems
│ ├── debates/ transcripts of multi-model discussions
│ └── validation/ perplexity novelty reports
├── research.md shared context across sessions (three-model share)
└── tools/ shell helpers (arxiv download, mineru, etc.)
All wiki content is LLM-compiled active knowledge, not a vector index. Re-compile when raw/ changes.
Phases
Phase 1 — Ingest
User: "ingest arXiv 2401.12345" or pastes a URL
Action:
1. Download PDF → use Bash: curl/arxiv-tool to idea-research/ideas/<slug>/raw/<arxiv-id>.pdf
2. Convert PDF → Markdown. Try mineru CLI first, marker as fallback:
mineru -p <pdf> -o <dir> (or: marker_single <pdf> <out>)
If neither installed, tell the user and skip — do not silently fall back to a text dump.
3. Save as raw/<arxiv-id>.md — preserve LaTeX for equations, HTML for tables.
4. Append a line to raw/INDEX.md with paper title, arxiv id, date ingested.
Never modify or delete files already in raw/. It is append-only.
Phase 2 — Compile Wiki
Trigger: "compile wiki" or after ingest of new papers
Action:
1. Read all raw/*.md.
2. For each new paper, write wiki/papers/<slug>.md with: one-paragraph TL;DR,
key claims, methods, datasets, results, limitations, related-work cross-refs as [[backlinks]].
3. Detect recurring concepts across papers; write or update wiki/concepts/<concept>.md
with cross-paper synthesis and [[backlinks]] to the paper notes.
4. Identify gaps and write wiki/gaps/<gap>.md — contradictions, missing ablations,
unverified claims. Each gap is a candidate idea seed.
5. Run lint checks: orphan links, contradictions across concept notes,
duplicate concept files with slightly different names.
The wiki is active knowledge: Claude writes it, Claude reads it. Do not treat as RAG.
Phase 3 — Obsidian
Pure consumption layer. Obsidian reads wiki/ via graph view + Dataview. Do not write to wiki/ from Obsidian side — Claude owns wiki/.
Phase 4 — Multi-Model Debate
Trigger: "debate <question>" or "讨论 <question>"
Precondition: wiki/ and debates/shared-research.md are up to date.
Action:
1. Prepare briefing: relevant wiki/papers + wiki/concepts + wiki/gaps + the question.
2. Invoke external models via Bash (installed CLIs only; if missing, skip that role and note it):
- gemini CLI → web-check role: search latest papers, verify facts
- codex CLI → adversarial role: find holes, counter-examples, prior art
- this Claude → synthesizer role: reconcile, align with wiki, arbitrate
3. Save transcript to debates/<date>-<slug>.md with each model's output labeled.
4. Extract disagreements; write them to wiki/gaps/ as new open questions.
Do not fabricate model outputs if a CLI is missing. Report "gemini CLI not installed, skipped."
Phase 5 — Novelty Validation
Trigger: "validate novelty of <idea>"
Action:
1. Load the idea text from ideas/<slug>/idea.md.
2. Call Perplexity (via WebFetch or curl to pplx-api if key is in env).
3. Cross-check with Semantic Scholar API for citation neighbourhood.
4. Write report to validation/<date>-novelty.md: prior-art matches with similarity notes,
open gaps, verdict (novel / incremental / covered).
Do not declare novelty without at least one external search step. If all external calls fail,
surface that clearly instead of guessing.
Phase 6 — Iterate
After validation, update:
- ideas/<slug>/idea.md with the verdict and next-step experiments
- wiki/gaps/ with new questions raised
- research.md with high-level direction change, if any
Trigger the next Phase-1 ingest when new seed papers are identified.
Invocation Patterns
When the user says things like:
- "帮我 ingest 这篇 arXiv" → Phase 1
- "compile wiki" / "编译知识库" → Phase 2
- "让三个模型讨论一下 " → Phase 4
- "validate 这个 idea 的新颖性" → Phase 5
- "开始新的 idea 研究: " → create ideas//, scaffold subdirs, ask user to paste seed papers
If the user gives a bare idea statement with no phase hint, default to Phase 5 (novelty validation) because it has the lowest cost and highest information gain early on.
External Tool Contract
Skill assumes these may or may not be installed. Detect, degrade gracefully, report honestly.
| Tool | Purpose | Detection |
|---|
mineru | PDF → Markdown | which mineru |
marker_single | PDF → Markdown (alt) | which marker_single |
gemini | Web-check role | which gemini |
codex | Adversarial role | which codex |
curl + PPLX_API_KEY env | novelty search | env var set |
Output Discipline
- Wiki files: one concept per file, YAML frontmatter with
type, tags, updated.
- Backlinks:
[[target]] or [[target|display]]; Obsidian parses both.
- No emoji unless the user added them. No "as an AI…" preambles. Short, dense notes.
- When the user says "update wiki after reading X", do not also rewrite unrelated files.
What This Skill Does Not Do
- Does not vector-index the wiki. Wiki is LLM-compiled active knowledge.
- Does not auto-run experiments. Experiment execution happens in the user's normal dev flow; the skill only organizes the knowledge around experiments.
- Does not push anything to git. The entire
idea-research/ tree is gitignored.
