메뉴
Evaluate the strength of evidence behind scientific claims based on study design, replication status, venue quality, sample size, and recency. Use when deciding how much weight to put on a finding, or when calibrating how confidently to write about a result.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Search HuggingFace Hub for datasets by keyword. Use when the user wants to find training data, benchmarks, or evaluation datasets for ML/NLP/CV research.
Fetch the FULL TEXT of an arXiv paper (all sections — introduction, method, results, conclusion). Use when you need to read beyond the abstract into the paper's actual content. Only works for arXiv papers. For metadata/abstract only, use paper-read. For local PDFs, use paper-read-pdf.
Two-layer memory system with grep-based recall for research sessions.
Track specific scientific claims across the literature over time — who made it, who replicated it, who challenged it, whether it still stands. Use when verifying a key assumption before building on it, or when checking whether a published result has been updated or superseded.
Scan papers for conflicting empirical claims, methodological disagreements, or opposing conclusions on the same topic. Use when writing discussion sections, evaluating conflicting results, or checking if a claim is contested before building on it.
Synthesize findings across multiple papers into a coherent narrative, structured comparison table, or temporal evolution. Use after collecting papers via survey or paper-search. Goes beyond summarizing individual papers to produce insights that only emerge when reading across the corpus as a whole.
| name | evidence-grading |
| description | Evaluate the strength of evidence behind scientific claims based on study design, replication status, venue quality, sample size, and recency. Use when deciding how much weight to put on a finding, or when calibrating how confidently to write about a result. |
| always | false |
Evidence is evaluated across 5 dimensions:
| Dimension | Strong | Weak |
|---|---|---|
| Study design | Controlled experiment, ablations, comparisons | Single-condition, no ablations |
| Replication | Replicated by independent groups | Single study, preprint only |
| Sample/scale | Large dataset, multiple seeds, statistical tests | Small dataset, single seed |
| Venue | Top-tier venue (NeurIPS/ICML/Nature/Science) | Workshop / preprint / self-published |
| Recency | Recent (last 2 years) | Older work that may have been superseded |
| Grade | Label | Criteria |
|---|---|---|
| A | Established | Multiple top-venue replications; widely accepted; in textbooks |
| B+ | Strong | Published top venue + ≥1 independent replication |
| B | Good | Top-venue paper, solid methodology, not yet replicated independently |
| C+ | Moderate | Smaller venue or preprint + single replication |
| C | Preliminary | Single top-venue paper, no independent replication |
| D | Weak | Workshop paper, preprint, limited evaluation |
| F | Insufficient | Single experiment, no baselines, no ablations |
| Grade | Suggested language |
|---|---|
| A | "It is established that..." / "[Method] has been shown to..." |
| B+ | "[Method] consistently demonstrates..." |
| B | "[Paper] shows that... (though independent replication is limited)" |
| C+ | "[Paper] suggests that... with moderate evidence" |
| C | "[Paper] provides preliminary evidence that..." |
| D | "[Paper] reports that... (preliminary; requires replication)" |
| F | Avoid citing as evidence; mention only for context |
paper-read to get: venue, year, citation countRead methodology (via paper-fetch or paper-read-pdf):
Use paper-search and claim-tracker:
paper-search --year-from [year]### Evidence Grade: [Claim]
- **Grade**: [A / B+ / B / C+ / C / D / F]
- **Source**: [Paper title] ([Year], [Venue])
- **Replication**: [Independent replications: N | Not yet replicated]
- **Study design**: [Controlled with ablations | Limited baselines | Single condition]
- **Scale**: [Large-scale, multi-seed | Small-scale, single seed]
- **Recency**: [Current | Potentially outdated — see [newer paper]]
- **Rationale**: [2–3 sentence explanation of the grade]
- **Recommended citation language**: "[Suggested phrasing]"
- **Caution**: [Any specific caveats about using this evidence]
claim-tracker — verify claim status first, then grade evidence strengthcontradiction-detection — contradicted claims automatically receive a grade penalty (max B-)reproduce skill — prioritize reproducing B or lower results that are being relied upon