paper-writer
VerifiedMedical/scientific paper writing workflow skill. Manages the full pipeline from literature search to submission-ready manuscript. Creates and manages a project directory with IMRAD-format section files, literature matrix, reference management, and quality checklists. Supports both English and Japanese papers. Triggers: 'write paper', 'paper-write', 'start manuscript', '論文を書く', '論文執筆', '論文プロジェクト', 'manuscript', 'research paper', '原稿作成'.
references/novelty-check.md
View on GitHub# Novelty Check — Phase −1 Guardrail
> **Placement:** Run once per surviving candidate question, after `templates/research-question.md`
> generates and scores candidates and **before** `templates/study-design.md` is opened.
> Output is the Novelty Assessment record (template at the end of this file).
>
> **Purpose:** The cheapest experiment is the one you don't run. Classify each gap
> honestly against the live literature. This is the direct implementation of the
> §5.2 guardrail in `references/ai-for-science-model.md`.
---
## 🤖 The Sakana v2 Anti-Pattern (Read This First)
Sakana AI Scientist v2 rated its own output "novel" when independent review found
the same territory already mapped. The failure was structural: the system searched
only what it already knew, accepted its own framing of the gap, and had no
adversarial step that could contradict it.
**The corrective is not asking AI to try harder. It is running a live query against
real literature databases the AI cannot hallucinate, then classifying the verdict
honestly — including the verdict "already answered, kill this."**
A human must read the top 5–10 closest papers and make the final call (💡 IDEA gate).
AI executes the sweep; the human owns the verdict.
---
## The Four Novelty Verdicts
| Verdict | Meaning | Decision |
|---------|---------|----------|
| **GENUINELY NOVEL** | No published answer exists; no ongoing trial/SR covers it | Proceed to study design |
| **INCREMENTAL** | Answered in a narrower population, older dataset, or lower-evidence design; your study adds meaningful precision | Proceed — but the differentiator must be explicit in the Introduction and pre-registration |
| **ALREADY ANSWERED** | A well-powered RCT, SR, or current clinical guideline settles the question | Kill or reframe. State *why* the existing evidence does not answer your specific version |
| **CONTESTED** | Conflicting high-quality evidence exists; the field has no consensus | Strong warrant to study — design to adjudicate, not to add a vote |
---
## Sweep Procedure
### Step 0 — Build the Query
Convert the candidate question from PICO elements (see `references/pubmed-query-builder.md`):
```
Population: [MeSH term] OR [free text synonym]
Intervention: [MeSH term] OR [free text synonym]
Comparison: (if applicable)
Outcome: [primary outcome term]
Combine: (P) AND (I) AND (O)
Add date filter: 2015[PDAT] : 3000[PDAT] ← last ~10 years as default
Add study-type filter if desired: systematic[sb] OR meta-analysis[pt]
```
Run the full combined query AND a loose version (drop the Outcome term) to
catch adjacent work. Write down the exact query string — it goes into the
Novelty Assessment record.
---
### Step 1 — PubMed MCP (Primary Biomedical Sweep)
PubMed is the primary database for clinical and biomedical questions.
Use the MCP tools directly:
**Initial search:**
```
mcp__claude_ai_PubMed__search_articles
query: "<your PICO query>"
max_results: 20
```
**Check each result's metadata:**
```
mcp__claude_ai_PubMed__get_article_metadata
pmid: "<PMID>"
```
**Expand from the closest paper (seed expansion):**
```
mcp__claude_ai_PubMed__find_related_articles
pmid: "<PMID of closest paper>"
max_results: 15
```
Seed expansion catches papers that share citation network topology but may not
share your exact terminology. Run it on the top 2–3 closest papers.
**Retrieve full text if key paper needs close reading:**
```
mcp__claude_ai_PubMed__get_full_text_article
pmid: "<PMID>"
```
**Convert IDs if you have a DOI but need a PMID:**
```
mcp__claude_ai_PubMed__convert_article_ids
ids: ["10.1001/..."]
id_type: "doi"
```
**Look up a specific citation by journal/year/author when suspected duplicate:**
```
mcp__claude_ai_PubMed__lookup_article_by_citation
journal: "Pediatrics"
year: 2022
first_author: "Smith"
```
Record: total hits, top 5–10 papers (PMID + title + year), and whether any
directly answer the question.
---
### Step 2 — Systematic Review & Protocol Registries
These catch work that would make your study redundant *before* it even publishes.
**Cochrane Library** — complete SRs:
```
https://www.cochranelibrary.com/search?q=<your+terms>
```
Fetch via WebFetch or `mcp__firecrawl__firecrawl_scrape`.
**PROSPERO** — registered SR protocols (including in-progress):
```
https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=...
https://www.crd.york.ac.uk/PROSPERO/search/SearchSummary.php?SuppliedID=<query>
```
If a PROSPERO record covers your exact question and is active, your independent
SR would be a duplicate — **kill or contact the team**.
**Epistemonikos** (SR synthesis across Cochrane + 50 databases):
```
https://www.epistemonikos.org/en/search?q=<query>
```
---
### Step 3 — Trial Registries (Ongoing Primary Research)
If an ongoing RCT or prospective study will answer the question in 2–3 years,
a new observational study may be pre-empted.
| Registry | URL | Scope |
|----------|-----|-------|
| ClinicalTrials.gov | `https://clinicaltrials.gov/search?term=<query>` | US + international |
| UMIN-CTR | `https://upload.umin.ac.jp/cgi-bin/ctr/ctr_view_j.cgi` | Japan |
| jRCT | `https://jrct.niph.go.jp/en-latest-detail/<id>` | Japan (new) |
| ISRCTN | `https://www.isrctn.com/search?q=<query>` | UK + global |
Filter for: Status = Recruiting OR Active, Not Recruiting. If a large RCT
has 3 years of follow-up left, plan to submit yours as a pilot/exploratory
framing, not a definitive trial.
---
### Step 4 — Broad Literature (OpenAlex, Europe PMC, Semantic Scholar)
For coverage beyond PubMed (non-indexed journals, preprints, citation graph):
**OpenAlex** (no API key required):
```
https://api.openalex.org/works?search=<url-encoded-query>&filter=publication_year:2015-&sort=cited_by_count:desc&per_page=20
```
Fetch via WebFetch. Sort by `cited_by_count:desc` to surface the most
influential work. A highly-cited paper directly answering your question is a
near-certain "ALREADY ANSWERED" signal.
**Europe PMC** (preprints + full-text search):
```
https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=<query>&format=json&resultType=core&pageSize=20
```
Catches medRxiv/bioRxiv preprints before PubMed indexes them.
**Semantic Scholar** (citation graph, influential-citation count):
```
https://api.semanticscholar.org/graph/v1/paper/search?query=<query>&fields=title,authors,year,citationCount,influentialCitationCount&limit=20
```
`influentialCitationCount` flags papers that methodologically shaped the field —
if one of these directly answers your question, that is a strong ALREADY ANSWERED.
**NotebookLM (optional — corpus grounding):**
If you have collected PDFs of relevant papers, use NotebookLM MCP
(`mcp__notebooklm-mcp__notebook_query`) to query across the local corpus.
Best for deep reading across 10–30 pre-loaded papers.
---
### Step 5 — Non-English & Adjacent Field Check
The honest-novelty checklist below governs this step. Minimum:
- Search with at least one non-English term if the question is clinically
active in Japan (e.g., UMIN-CTR search in Japanese characters).
- Run one adjacent-field query (e.g., if your question is pediatric,
run the adult version to see if the adult answer translates).
---
## Tools Summary
| Tool | Endpoint / MCP | Best For |
|------|---------------|---------|
| PubMed MCP | `mcp__claude_ai_PubMed__search_articles` | Primary biomedical sweep |
| PubMed related papers | `mcp__claude_ai_PubMed__find_related_articles` | Citation-network expansion |
| PubMed full text | `mcp__claude_ai_PubMed__get_full_text_article` | Close reading of key papers |
| PubMed metadata | `mcp__claude_ai_PubMed__get_article_metadata` | Confirm study type, date, journal |
| PubMed ID convert | `mcp__claude_ai_PubMed__convert_article_ids` | DOI ↔ PMID reconciliation |
| PubMed citation lookup | `mcp__claude_ai_PubMed__lookup_article_by_citation` | Check suspected duplicate |
| OpenAlex API | `https://api.openalex.org/works?search=…` | Broad coverage, citation count sort |
| Europe PMC API | `https://www.ebi.ac.uk/europepmc/…` | Preprints + full-text search |
| Semantic Scholar API | `https://api.semanticscholar.org/graph/v1/paper/search?…` | Citation graph, influential papers |
| Cochrane Library | `https://www.cochranelibrary.com/search?q=…` | Completed systematic reviews |
| PROSPERO | `https://www.crd.york.ac.uk/PROSPERO/…` | In-progress SR registrations |
| Epistemonikos | `https://www.epistemonikos.org/en/search?q=…` | SR synthesis across 50 databases |
| ClinicalTrials.gov | `https://clinicaltrials.gov/search?term=…` | Ongoing trials (US + international) |
| UMIN-CTR | `https://upload.umin.ac.jp/cgi-bin/ctr/…` | Ongoing trials (Japan) |
| jRCT | `https://jrct.niph.go.jp/…` | Ongoing trials (Japan, new system) |
| NotebookLM MCP | `mcp__notebooklm-mcp__notebook_query` | Local PDF corpus grounding |
Fetch non-MCP endpoints via WebFetch, `mcp__firecrawl__firecrawl_scrape`, or
`mcp__tavily__tavily_search` as available.
---
## Honest-Novelty Checklist
Before assigning a GENUINELY NOVEL verdict, confirm you can check every box:
- [ ] Searched with ≥2 synonym/alternative term sets (not just your preferred phrasing)
- [ ] Searched non-English sources if clinically active in Japan (UMIN-CTR; J-STAGE via WebFetch)
- [ ] Searched preprint servers (Europe PMC catches medRxiv/bioRxiv)
- [ ] Ran seed expansion (`find_related_articles`) on the top 2–3 closest papers
- [ ] Checked for an ongoing trial that will answer the same question (ClinicalTrials.gov, UMIN, jRCT)
- [ ] Checked PROSPERO for a registered SR in progress
- [ ] Ran the adjacent-field version of the query (e.g., adult analogue of a pediatric question)
- [ ] Sorted OpenAlex by `cited_by_count:desc` — no high-citation paper directly answers this
- [ ] The "gap" is a genuine absence of evidence, not an unsearched corner of the literature
- [ ] A clinician-expert reviewed the top papers and agrees the question is unanswered
If any box is unchecked, the verdict cannot be GENUINELY NOVEL.
---
## Decision Tree
```
Candidate research question
│
▼
Run PubMed MCP sweep (Steps 1–2)
│
├── Cochrane SR / PROSPERO active review already covers this?
│ YES ──► ALREADY ANSWERED → Kill or contact SR team
│ NO ──► continue
│
▼
Top 5–10 papers found. Does any directly answer your primary endpoint
in your target population at adequate power?
│
├── YES, one or more well-powered RCT/SR/guideline
│ ──► ALREADY ANSWERED → Kill
│ OR reframe: "why doesn't the existing answer
│ apply to our population/setting/time?" — if
│ compelling, relabel INCREMENTAL and proceed
│
├── YES, but lower-quality evidence only (observational,
│ small N, older dataset, different subgroup)
│ ──► INCREMENTAL → Proceed, explicit differentiator required
│
├── YES, and conflicting results across studies
│ ──► CONTESTED → Proceed, design to adjudicate
│
└── NO papers close to the question
│
▼
Honest-novelty checklist — all 10 boxes checked?
│
├── NO ──► Complete missing steps, re-run sweep
└── YES ──► GENUINELY NOVEL → Proceed to study design
```
---
## Novelty Assessment Record (Output Template)
Fill one record per candidate question that passes initial FINER scoring.
```markdown
## Novelty Assessment
**Date:** YYYY-MM-DD
**Question:** [Exact PICO-formatted question from research-question.md]
**Assessed by:** [Name] (human) + AI sweep
### Query Used
- PubMed string: `[exact query]`
- OpenAlex string: `[exact query]`
- Date bounds: [YYYY]–present
- Synonyms tested: [list]
### Registry Check
- PROSPERO: [active protocol? Y/N — title if Y]
- ClinicalTrials.gov: [ongoing trial? Y/N — NCT if Y]
- UMIN/jRCT: [ongoing trial? Y/N — ID if Y]
### Top Existing Papers (5–10 closest)
| # | PMID / DOI | Title (short) | Year | Design | N | Verdict relevance |
|---|-----------|---------------|------|--------|---|------------------|
| 1 | | | | | | |
| 2 | | | | | | |
| … | | | | | | |
### What They Leave Open
[2–4 sentences: what the existing literature does NOT answer that this question does]
### Honest-Novelty Checklist
- [ ] ≥2 synonym sets searched
- [ ] Non-English / J-STAGE searched (if applicable)
- [ ] Preprints checked (Europe PMC)
- [ ] Seed expansion run on top 2–3 papers
- [ ] Ongoing trials checked
- [ ] PROSPERO checked
- [ ] Adjacent-field query run
- [ ] OpenAlex sorted by citation count — no high-citation direct answer
- [ ] "Gap" confirmed as genuine, not just unsearched
- [ ] Clinician-expert sign-off
### Verdict
**[ ] GENUINELY NOVEL / [ ] INCREMENTAL / [ ] ALREADY ANSWERED / [ ] CONTESTED**
### Differentiator Statement
"Our study differs from the existing literature because…"
[1–2 sentences. This language goes directly into the Introduction.]
### Decision
**[ ] Proceed to study-design.md**
**[ ] Reframe — revised question:** [new question]
**[ ] Kill — reason:** [reason]
**Approved by (💡 IDEA gate):** ___________ Date: ___________
```
---
## Cross-References
- `references/ai-for-science-model.md` §5.2 — the guardrail this file implements
- `templates/research-question.md` — upstream: generates candidate questions
- `references/pubmed-query-builder.md` — how to build the PICO query in Step 0
- `templates/study-design.md` — downstream: opens only after NOVEL/INCREMENTAL/CONTESTED verdict
- `templates/preregistration.md` — the downstream lock that freezes the plan
- `references/adversarial-review.md` — the downstream red-team that stress-tests the claim
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |