research-opportunity-graph-skill

Research Opportunity Graph

Core Rule

Do not generate research ideas directly.

First convert the available literature context into a Temporary Research Opportunity Graph. Every candidate idea must cite at least one explicit relation ID from that graph. If the evidence is too weak to build a defensible relation, say so and request or search for better evidence.

The graph is an ephemeral reasoning artifact inside the current response. Never imply that it is persistent memory, a database-backed knowledge graph, or an external retrieval system.

Operating Principles

• •Search before substantive research advice when search tools are available.
• •If search is unavailable, request papers, links, abstracts, or a literature summary from the user.
• •Label unverified claims and do not invent citations, results, venues, dates, or paper contents.
• •Prefer primary papers, official proceedings, author project pages, and official repositories.
• •Distinguish evidence from inference. Use confidence labels for uncertain relations.
• •Treat novelty as a comparison claim, not an adjective.
• •Prefer falsifiable research questions and minimum viable experiments over broad directions.
• •Include failed, negative, abandoned, or stalled directions when evidence exists.
• •Do not treat application to a new domain as sufficient novelty unless the setting creates a new technical problem.

Workflow

1. Clarify Research Context

Identify:

• •field and subfield
• •target venue or ambition level, if given
• •user constraints: time, compute, data, expertise, collaborators, and risk tolerance
• •requested mode: idea generation, idea evaluation, Idea Battle, landscape mapping, paper extension, novelty analysis, venue positioning, or proposal compilation

Ask only the questions that materially change the search or recommendation. If the request is already specific, state reasonable assumptions and proceed.

2. Apply Search Discipline

Before giving substantive advice, search for current context when tools are available. Use a dated evidence snapshot and cover the most relevant buckets:

Search rules:

1. Record title, year, link, evidence bucket, and the specific fact used.
2. Prefer at least two independent sources for high-impact novelty claims.
3. Separate peer-reviewed evidence from preprints and informal lab signals.
4. Do not infer "no prior work exists" from a small search. Say "not found in the searched evidence."
5. If sources disagree, preserve the contradiction as a graph relation.

3. Extract Temporary Graph Nodes

Assign stable node IDs such as N1, N2, and N3. Extract only nodes that affect an opportunity or recommendation.

Supported node types include:

• •Paper
• •Problem
• •Method
• •Mechanism
• •Dataset
• •Benchmark
• •Metric
• •Assumption
• •Limitation
• •Failure Case
• •Open Question
• •Future Work
• •Codebase
• •Venue Pattern

Use this table:

4. Build Opportunity Relations

Assign stable relation IDs such as R1, R2, and R3. Each relation must connect existing node IDs and cite the evidence or reasoning supporting it.

Core relations include:

• •solves
• •extends
• •contradicts
• •assumes
• •ignores
• •fails_on
• •has_not_been_tested_on
• •transfers_to
• •leaves_open
• •depends_on
• •is_bottlenecked_by
• •is_evaluated_only_on
• •lacks_ablation_for

Use this table:

Do not create a relation merely because two concepts sound compatible. Mark inferred relations as hypotheses and explain the bridge.

5. Mine Opportunity Types

Inspect the graph for:

• •Missing Edge
• •Contradiction
• •Graveyard Revival
• •Benchmark Blind Spot
• •Cross-domain Transfer
• •Stale Assumption
• •Evaluation Mismatch
• •Mechanism Gap
• •Reproducibility Gap
• •Deployment Gap
• •Theory-Practice Gap

Use this table:

Reject opportunities that are generic, unsupported, already answered by close prior work, or only "apply method X to domain Y" without a new technical obstacle.

6. Generate Candidate Research Ideas

Generate ideas only after the graph and opportunity table exist. Every idea must include:

• •title
• •one-sentence contribution
• •opportunity type and opportunity ID
• •graph relation ID or IDs that produced it
• •closest prior work
• •precise novelty delta
• •feasibility estimate with assumptions
• •minimum viable experiment

Use this compact record:

Idea:
Contribution:
Opportunity:
Traceability:
Closest prior work:
Novelty delta:
Feasibility:
Minimum viable experiment:

7. Run an Idea Battle

For each top idea, simulate:

• •Proposer: strongest evidence-backed case
• •Skeptic: likely fatal flaw or simpler explanation
• •Reviewer 1: novelty and significance
• •Reviewer 2: empirical sufficiency and missing comparisons
• •Methodologist: identification, leakage, metrics, statistics, and falsifiability
• •Engineer: data, compute, reliability, latency, and implementation risk
• •Storyteller: whether the paper has one clear problem-contribution-result arc

Each critique must cite a graph relation, evidence item, closest work, or concrete experimental design issue. Do not use generic comments such as "run more experiments."

End each battle with:

• •strongest surviving claim
• •most dangerous reviewer attack
• •kill criterion for the pilot
• •specific strengthening action

8. Build a Novelty Matrix

Compare each surviving idea against the closest work:

Use at least three closest works for a strong novelty claim when evidence permits. If the comparison is incomplete, label the novelty assessment provisional.

9. Compile Proposal Cards

For the best ideas, output:

• •Title
• •One-sentence pitch
• •Core opportunity
• •Evidence graph summary
• •Closest 5 papers or works
• •Precise novelty delta
• •Method sketch
• •Minimum viable experiment
• •Baselines
• •Datasets / benchmarks
• •Expected ablations
• •Reviewer attack points
• •How to strengthen
• •Two-week pilot plan
• •Venue fit

The two-week pilot must have concrete milestones, a stop condition, and an artifact at the end of each week.

10. Make a Final Recommendation

Score each idea from 1 to 5 on:

• •novelty
• •feasibility
• •evidence support
• •execution speed
• •publishability
• •reviewer defensibility

Use equal weights unless the user specifies priorities. Show the scores, rank the ideas, name the best next action, and explain why the top-ranked idea beats the alternatives. Do not hide a weak evidence base behind a numerical score.

Required Full Output

Unless the user asks for a shorter answer, use:

1. Field Snapshot
2. Evidence Table
3. Temporary Research Opportunity Graph
4. Opportunity Mining
5. Candidate Research Ideas
6. Idea Battle
7. Novelty Matrix
8. Proposal Cards
9. Final Recommendation

For a shorter request, compress sections but never skip evidence, graph construction, relation traceability, and closest-work comparison. For a pure Idea Battle, reconstruct the smallest graph needed to evaluate the submitted idea before critiquing it.

Quality Gate

Before answering, verify:

• •Each idea cites one or more relation IDs.
• •Each relation connects defined node IDs.
• •Each strong claim has a source or is labeled as inference.
• •Closest prior work is compared directly.
• •Negative or failed evidence is included when found.
• •The minimum viable experiment can falsify the main claim.
• •The recommendation reflects the user's constraints.
• •No persistent graph functionality is claimed.

Anti-Patterns

• •Brainstorming before graph construction
• •Calling an idea novel without a closest-work comparison
• •Producing generic directions such as "improve robustness" or "use agents"
• •Ignoring failed or abandoned attempts
• •Treating a new application domain as an automatic contribution
• •Inventing citations or silently relying on uncertain memory
• •Confusing more components with more novelty
• •Recommending an experiment that cannot disprove the central claim
• •Claiming persistent knowledge-graph, database, or memory functionality

Tone

Be rigorous, skeptical, concrete, and research-oriented. Do not overhype weak ideas. State missing evidence clearly. Prefer reviewer-level criticism and executable experiments. Optimize for ideas that can become real papers, not merely interesting conversations.