novelty-check — скилл для Claude Code: что делает и как установить

---
name: novelty-check
description: Verify research idea novelty against recent literature. Use when user says "查新", "novelty check", "有没有人做过", "check novelty", or wants to verify a research idea is novel before implementing.
argument-hint: "[method-or-idea-description]"
allowed-tools: Bash(*), WebSearch, WebFetch, Grep, Read, Glob, mcp__codex__codex, mcp__antigravity__antigravity, mcp__antigravity__antigravity-reply, mcp__grok__grok, mcp__grok__grok-reply, mcp__claude-review__review, mcp__claude-review__review_reply, mcp__claude-review__review_start, mcp__claude-review__review_reply_start, mcp__claude-review__review_status
---

# Novelty Check Skill

> **Before each reviewer slot**, follow [Selectable Reviewers](../shared-references/reviewer-backends.md):
> resolve invocation → `.aris/reviewer.json` → Claude Code default (Codex MCP)
> with `reviewer_route.py` through the canonical helper resolver, preserving this
> slot's tier, fresh contexts, source/image inputs and deterministic gates.
> Codex examples/pins below apply only to `backend=codex`; other routes use their
> own first-call/reply contract. Save the route and actual returned identity;
> unavailable selections are `REVIEW_UNAVAILABLE`. Classify acceptance from actual
> families/completion/scope. Forward invocation reviewer/model/effort overrides
> to reviewer-bearing children. Generation workers are unchanged.


Check whether a proposed method/idea has already been done in the literature: **$ARGUMENTS**

## Constants

- REVIEWER_MODEL = `gpt-6-astra` — Model used via Codex MCP. For `backend=codex`, use an OpenAI model (e.g., `gpt-6-astra`, `gpt-6.1-sol`)

## Instructions

Given a method description, systematically verify its novelty:

### Phase A: Extract Key Claims
1. Read the user's method description
2. Identify 3-5 core technical claims that carry the claimed delta:
   - What is the method?
   - What problem does it solve?
   - What is the mechanism?
   - What makes it different from obvious baselines?

### Phase B: Multi-Source Literature Search
For EACH core claim, search using ALL available sources:

1. **Web Search** (via `WebSearch`):
   - Search arXiv, Google Scholar, Semantic Scholar
   - Use specific technical terms from the claim
   - Try at least 3 different query formulations per claim
   - Include year filters for 2024-2026

2. **Known paper databases**: Check against:
   - ICLR 2025/2026, NeurIPS 2025, ICML 2025/2026
   - Recent arXiv preprints (2025-2026)

3. **Read abstracts**: For each potentially overlapping paper, WebFetch its abstract and related work section

### Phase C: Cross-Model Verification
Call REVIEWER_MODEL via Codex MCP (`mcp__codex__codex`) with xhigh reasoning.
When the method description plus the Phase-B paper list is more than a short
note, avoid pasting it inline into the MCP prompt. Write a dossier file such as
`NOVELTY_DOSSIER.md` (or a project-local equivalent) containing the method
description, core claims, candidate papers, and the exact questions below, then
send only the file path:
> Resolve this reviewer slot first; the following is the `backend=codex` example.

```
mcp__codex__codex:
  model: gpt-6-astra
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    Read the novelty dossier at <absolute path to NOVELTY_DOSSIER.md> and
    follow all instructions in it.
```
Dossier contents should include:
- The proposed method description
- All papers found in Phase B
- Ask: "Is this method novel? What is the closest prior work? What is the delta?"
- The NOVELTY VERDICT LIMITS block below, verbatim — the reviewer judges under it

### The verdict limits

Copy this block **verbatim** into the reviewer's briefing; the report in
Phase D is judged under it too.

```
=== NOVELTY VERDICT LIMITS (these bound how you judge, never how widely you search) ===
Search exhaustively; judge calibrated. Two failures waste months equally:
passing an idea a published paper already contains, and killing a viable idea
because the territory has neighbors.
1. Proximity is information, not a verdict. Someone working nearby goes in the
   report; it is not by itself a reason to reject.
2. ABANDON has exactly one qualification: a specific published paper already
   contains this result — name that paper. No named paper, no ABANDON.
3. Crowded-but-deltaed is PROCEED: state the delta in one sentence a reviewer
   could verify. Thin or contested delta is PROCEED WITH CAUTION — say what
   would make it carry, not why it should die. CAUTION is not a safe middle:
   if you cannot name the specific thing that makes the delta thin, the
   verdict is PROCEED.
4. Concurrent or competing work is not a veto. That is a race — report it and
   let the user decide whether to run it.
5. A direct attack on a central problem is legitimate novelty when nobody has
   executed it well. "This area is hot" does not mean "this area is taken."
6. This check is an early gate, never the last one — more triage, pilots, or
   external review still stand between any idea and a paper, whatever order
   this run uses. A wrongly passed idea dies cheaply at one of them; a wrongly
   killed idea is never seen again. When torn between two verdicts, choose the
   more permissive one.
Say plainly when an idea clears the check. Do not manufacture overlap.
```

### Phase D: Novelty Report
Output a structured report:

```markdown
## Novelty Check Report

### Proposed Method
[1-2 sentence description]

### Core Claims
1. [Claim 1] — Closest: [paper] — What stays unknown or different: [delta]
2. [Claim 2] — Closest: [paper] — What stays unknown or different: [delta]
...

### Closest Prior Work
| Paper | Year | Venue | Overlap | Key Difference |
|-------|------|-------|---------|----------------|

### Overall Novelty Assessment
- Score: X/10 (anchor: 5/10 = has clear neighbors but a defensible delta worth
  a pilot; reserve 1-3 for results a named published paper already contains)
- Recommendation: PROCEED / PROCEED WITH CAUTION / ABANDON (per the verdict
  limits: crowded-but-deltaed ground is PROCEED; ABANDON must name the paper)
- Key differentiator: [what makes this unique, if anything]
- Risk: [what a reviewer would cite as prior work]

### Suggested Positioning
[State the delta honestly in one sentence a reviewer could verify]
```

### Important Rules
- Two failures waste months equally: a false novelty claim, and a viable idea
  abandoned because the territory has neighbors. Be brutally honest in both
  directions — and when an idea clears the check, say so plainly.
- Novelty can live in the combination or the finding even when every
  individual claim rates LOW — judge the idea, not each claim in isolation.
  Known parts arranged to reveal something unknown are novel.
- "Applying X to Y" earns novelty by what the application reveals — a
  non-obvious interaction, failure mode, or insight. Judge the revelation, not
  the template.
- Check both the method AND the experimental setting for novelty
- If the method is not novel but the FINDING would be, say so explicitly
- Always check the most recent 6 months of arXiv — the field moves fast
- **Anti-hallucination for Closest Prior Work.** Every paper in the prior-work table must pass pre-search verification via `verify_papers.py` (canonical name resolved per [`shared-references/integration-contract.md`](../shared-references/integration-contract.md) §2; 3-layer arXiv / CrossRef / Semantic Scholar fallback inside the helper itself). Policy D1 (primary + degraded-output fallback): if the helper is unresolved **or** its invocation fails, tag candidate entries `[UNVERIFIED]` and surface the uncertainty rather than dropping them. Never fabricate arXiv IDs, DOIs, or titles from memory. Full protocol in [`shared-references/citation-discipline.md`](../shared-references/citation-discipline.md) § Pre-Search Verification Protocol.

## Review Tracing

After each `mcp__codex__codex` or `mcp__codex__codex-reply` reviewer call, save the trace following `shared-references/review-tracing.md` (Policy C — forensic; never silently skip). Use `save_trace.sh` (resolved per the chain in `shared-references/integration-contract.md` §2) or write files directly to `.aris/traces/<skill>/<date>_run<NN>/`. Respect the `--- trace:` parameter (default: `full`).