You don’t need another paper list. You need to know what transfers. Use this agent SKILL to : Find papers similar to your work. Then expose the main difference. To see how their research transfers.
Most literature searches stop at relevance. Sim-Diff asks what a paper can contribute to your work, by first scraping literature that share similarity with your work: be it pursuing similar objectives or using similar techniques, and then exposing where exactly do they differ from your work.
This becomes important, since matching based on titles, keywords, or even the abstract does not always reveal full informations about the work. Often papers uses techniques that are of your interest that remains unhighlighted in the abstract, this skill explicitly asks the agent to treat research objectives pursued and techniques used separately for a better relevance.
| Step | Question |
|---|---|
| SIM | What objective, technique, mechanism, or structure is genuinely shared? |
| DIFF | What changes in the assumptions, representation, scope, or guarantee? |
| TRANSFER | What lemma, construction, algorithm, counterexample, or experiment should be tested next? |
This matters because the closest topic may offer nothing reusable, while a distant field may contain exactly the mechanism you need. The vector geometry is a reasoning metaphor—not a claimed embedding system, cosine-search engine, or novelty certificate.
Give your agent the Sim-Diff skill, your research question, and either a selected paper set, research notes or an authorized search scope.
Use $sim-diff to compare my research question with the supplied papers.
Separate objective and technique similarity.
Return each useful neighbor as SIM → DIFF → TRANSFER.
Cite the evidence and mark unresolved claims.
New-paper discovery requires authorized browsing. Comparing a supplied corpus does not.
USER's notes / explicit question
│
▼
@frozen-anchor ◀── #BuildAnchor(@notes, @question)
│
┌──────────────────────────────────────────┼──────────────────────────────────────────┐
│ │ │
▼ ▼ ▼
#ObjectiveSearch(@frozen-anchor) #TechniqueSearch(@frozen-anchor) #BridgeSearch(@frozen-anchor)
│ │ │
└──────────────────────────────────────────┼──────────────────────────────────────────┘
│
▼
#InspectPrimaryEvidence(@candidates)
│
▼
@two-axis-judgments
│
▼
#SIM(@evidence) ──▶ #DIFF(@similarities) ──▶ #TRANSFER(@differences)
│
▼
@report and @receipts
│
▼
#ReviewAndPrune(@report)
- A balanced, domain-agnostic Sim-Diff skill.
- Experimental lite, geometry, and evidence-first variants.
- An optional standard-library experimental lab for frozen comparison packets, structural validation, and narrow fixture metrics.
- Reproducibility, provenance, and validation records.
The package passed 38 deterministic tests covering its structure, fixtures, snapshots, and helper boundaries. These checks do not establish model quality, scientific correctness, exhaustive literature coverage, host isolation, or superiority of one variant.
See VALIDATION.md for the tested boundary and START_HERE.md for the complete local workflow.
MIT © 2026 Rajarsi Pal.