novelty-check - SKILL.md Agent Skill

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." allowed-tools: Bash(*), WebSearch, WebFetch, Grep, Read, Glob

Override for Codex users who want Claude Code CLI, not a second Codex agent, to act as the reviewer/helper. Install this package after skills/skills-codex/*.

Whenever the upstream skill asks for an external reviewer/helper, write the complete focused prompt to $PROMPT_FILE. For a one-shot independent review, run:

claude -p --dangerously-skip-permissions --output-format json --model opus --effort max < "$PROMPT_FILE" | tee "$RAW_REVIEW_JSON"

For multi-round reviewer discussion, keep automation non-interactive but preserve continuity with --session-id on the first call and --resume on follow-up calls; see ../shared-references/claude-cli-review.md.

Novelty Check Skill

Classify novelty risk and positioning routes for: $ARGUMENTS

Constants

REVIEWER_MODEL = claude-cli — Claude reviewer invoked through direct claude -p CLI calls, using --session-id / --resume for multi-round discussion, following ../shared-references/claude-cli-review.md.
NOVELTY_POLICY = positioning-first — Load ../shared-references/research-posture.md before judging novelty. Similar work is not automatically fatal.
CONCURRENT_WORK_WINDOW = 3 months — Recent work goes to orbit-research/CONCURRENT_WORK_WATCHLIST.md by default.

Instructions

Given a method description, systematically classify its novelty posture:

Phase A: Extract Key Claims

Read the user's method description
Identify 3-5 central factual/method/benchmark/paper-bearing claims whose novelty matters:
- What is the method?
- What problem does it solve?
- What is the mechanism?
- What makes it different from obvious baselines?
- What contribution type is being attempted: normal method, empirical finding, benchmark/baseline, reproduction-plus, system, focused mechanism, or breakthrough?

Phase B: Multi-Source Literature Search

For EACH core claim, search using ALL available sources:

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
Known paper databases: Check against:
- ICLR 2025/2026, NeurIPS 2025, ICML 2025/2026
- Recent arXiv preprints (2025-2026)
Read abstracts: For each potentially overlapping paper, WebFetch its abstract and related work section

Phase C: Cross-Model Positioning Review

Call REVIEWER_MODEL via Claude CLI reviewer (claude -p) with xhigh reasoning:

config: {"Claude CLI `--effort max`": "xhigh"}

Prompt should include:

The proposed method description
All papers found in Phase B
The default posture from ../shared-references/research-posture.md
Ask: "Classify novelty posture using CLEAR_SPACE / RELATED_BUT_DIFFERENT / CONCURRENT_WORK / WEAK_BLOCKER / STRONG_BLOCKER / POSITIONING_TARGET / REPRODUCTION_TARGET. What is the closest prior work, what overlaps, how reliable is it, and what survival/positioning route remains?"

Phase D: Positioning Report

Output a structured report:

## Novelty Positioning Report

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

### Core Claims
1. [Claim 1] — Novelty Posture: [class] — Closest: [paper] — Survival route: [route]
2. [Claim 2] — Novelty Posture: [class] — Closest: [paper] — Survival route: [route]
...

### Closest Prior Work
| Paper | Date | Venue / Status | Overlap Dimensions | Reliability | Code/Data | Reproducible? | Novelty Posture |
|-------|------|----------------|--------------------|-------------|-----------|---------------|-----------------|

### Overall Novelty Posture
- Score: X/10
- Novelty Posture: CLEAR_SPACE / RELATED_BUT_DIFFERENT / CONCURRENT_WORK / WEAK_BLOCKER / STRONG_BLOCKER / POSITIONING_TARGET / REPRODUCTION_TARGET
- Blocker strength: NONE / WEAK / STRONG
- Strong blocker criteria met: YES / NO
- Key differentiator: [what makes this unique, if anything]
- Risk: [what a reviewer would cite as prior work]
- Prior-work reliability: peer-reviewed / arXiv-only / recent concurrent / unpublished / partially open
- Public detail: sufficient / partial / missing
- Code/data availability: available / partial / unavailable

### Positioning Route
[How to frame the contribution without overclaiming]

### Survival Route
[For every non-strong-blocker, at least one viable route: narrower regime, stronger
evidence, reproduction-plus, benchmark/baseline, system, mechanism analysis, or
conditional empirical finding]

### Concurrent Work Watchlist
[If any work is from the last 3 months, add or recommend adding it to
orbit-research/CONCURRENT_WORK_WATCHLIST.md. Do not rewrite the proposal unless it is a
STRONG_BLOCKER.]

Important Rules

Be honest but opportunity-preserving. Your job is to classify novelty risk and find viable positioning, not to kill promising ideas prematurely.
Do not recommend abandonment unless the STRONG_BLOCKER criteria in ../shared-references/research-posture.md are met.
For every non-strong-blocker, propose at least one positioning strategy.
"Applying X to Y" can be a valid normal-paper route when the setting, evidence, control design, or finding is interesting and honest.
Check both the method AND the experimental setting for novelty
If the method overlaps prior work but the FINDING would be differentiated, say so explicitly
Always check the most recent 6 months of arXiv; treat work from the last 3 months as concurrent by default and add it to the watchlist
After STOP A, ordinary new related work should not destabilize a frozen proposal. Use the watchlist unless the work is a STRONG_BLOCKER.

Review Tracing

After each claude -p or claude -p --resume reviewer call, save the trace following ../shared-references/review-tracing.md. Resolve save_trace.sh via that shared resolver, or write files directly to .aris/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).