Research Opportunity Graph
DajunG-77/research-opportunity-graph-skill
Ideação de pesquisa estruturada a partir de evidências literárias
Ver no GitHubUma skill prompt-only que transforma literatura atual em ideias de pesquisa fundamentadas em evidências e planos de experimento práticos. Constrói um mapa temporário do conhecimento existente, lacunas de pesquisa, e estratégias de teste, mantendo o raciocínio ancorado em contexto acadêmico real.
Instalação
npx convez-ai DajunG-77/research-opportunity-graph-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar ideias de pesquisa conectadas a trabalhos existentes
- Avaliar ideias de pesquisa contra estado da arte
- Desenhar experimentos realistas e testáveis
- Identificar lacunas genuínas em campos de pesquisa
Veredito

Skill bem estruturada para pesquisadores que precisam gerar ideias ancoradas em evidência, não apenas brainstorming genérico. Implementação prompt-only é portável, mas requer usuário com maturidade acadêmica para extrair máximo valor da abordagem rigorosa.
README
A tradução para o português deste README ainda está na fila.
Research Opportunity Graph
A prompt-only research skill that turns current literature into clear, evidence-backed research ideas and practical experiment plans.
Most research prompts jump from search results to brainstorming. This skill first builds a temporary map of:
what researchers already know -> what is still missing -> why the missing piece matters -> how to test it
The map is used internally to keep the reasoning grounded. Users receive a plain-English explanation rather than graph IDs, relation codes, or database terminology.
What Makes It Different?
Typical research ideation prompts:
- search for papers
- summarize them
- produce a list of possible ideas
- give generic criticism
Research Opportunity Graph follows a stricter workflow:
Search -> Understand the evidence -> Connect findings and limitations -> Find research gaps -> Design experiments -> Stress-test the best ideas
The goal is not to produce more ideas. The goal is to produce fewer ideas that are:
- clearly connected to existing evidence
- different from the closest prior work
- specific enough to test
- honest about novelty risk
- realistic under the user's time, data, and compute constraints
What Is a Temporary Opportunity Graph?
It is simply a structured reasoning map created for the current task.
For example:
Existing work:
Several systems let an AI agent revise a forecast using retrieved context.
Known weakness:
Irrelevant or misleading context can make a good forecast worse.
Missing piece:
The agent has no reliable way to decide whether a revision should be accepted.
Research opportunity:
Build and test an evidence gate that accepts only revisions likely to improve
forecast accuracy and calibration.
This is not a persistent knowledge graph. There is no database, vector store, API integration, or memory backend.
Install
Claude Code
git clone https://github.com/DajunG-77/research-opportunity-graph-skill.git \
~/.claude/skills/research-opportunity-graph-skill
Restart Claude Code, then ask for research idea generation, idea evaluation, reviewer-style criticism, or a paper plan.
Codex
git clone https://github.com/DajunG-77/research-opportunity-graph-skill.git \
"${CODEX_HOME:-$HOME/.codex}/skills/research-opportunity-graph-skill"
The repository can also be used with other agents that support Markdown skill instructions.
Example Prompts
Find three realistic research opportunities in LLM agents for time-series
forecasting. Explain the field first and assume I am new to the topic.
Find research opportunities in using multimodal LLMs to reason about and
correct compositional errors in text-to-image generation.
Stress-test this idea for NeurIPS: adaptive test-time compute for forecasting
agents. Tell me what is already known, what is actually new, and what experiment
would kill the idea.
Turn this accepted paper into two substantial extension ideas. Avoid minor
benchmark additions and explain the novelty in simple language.
Example Output Preview
1. Plain-English Summary
Current forecasting agents can retrieve external context and revise a model's prediction. The main open problem is not whether an agent can revise a forecast, but whether it can recognize when the new context is trustworthy enough to justify the revision.
2. What Existing Research Has Done
| Work | What it contributed | What it did not fully answer |
|---|---|---|
| Context-aware forecasting agent | Uses retrieved events to adjust predictions | When should the original forecast be kept? |
| Probabilistic forecasting model | Produces uncertainty-aware predictions | How should an agent use that uncertainty to choose actions? |
| Agent evaluation benchmark | Tests multi-step reasoning | How often does additional context make the answer worse? |
3. Research Opportunity Map
| What we know | What is missing | Why it matters |
|---|---|---|
| Agents can revise forecasts using external information | A reliable accept/reject rule for each revision | A fluent explanation can hide a worse prediction |
| Forecasting models can estimate uncertainty | A test of whether uncertainty predicts when agent tools help | Always using the agent wastes cost and may damage calibration |
4. Research Idea
Title: Evidence-Gated Forecast Revision
Research question: Can an agent test whether newly retrieved context is useful before changing a strong baseline forecast?
Why this idea exists: Existing work shows that contextual revision is possible, while evaluation work shows continuing uncertainty and reasoning failures. The missing step is a decision rule that protects the original forecast from harmful revisions.
What is new: The contribution is not another forecasting agent. It is an explicit accept, reject, or abstain decision for every proposed revision, together with a benchmark that measures harmful revisions.
Minimum experiment:
- Use two public forecasting datasets with known event information.
- Compare no revision, always revise, a simple confidence threshold, the proposed evidence gate, and an oracle.
- Measure forecast error, calibration, harmful-revision rate, latency, and model calls.
- Reject the idea if the evidence gate cannot beat the best simple threshold at equal cost.
5. Reviewer Stress Test
| Question | Honest answer |
|---|---|
| Why might this be accepted? | It studies a concrete failure introduced by agentic forecasting. |
| Why might it be rejected? | Reviewers may see it as a complicated retrieval filter. |
| What comparison is essential? | A strong similarity or confidence threshold under the same cost. |
| What result would kill the paper? | The proposed gate does not generalize across datasets or backbones. |
6. What Is Actually New?
| Closest work already does | This project adds | Novelty risk |
|---|---|---|
| Retrieves context and revises forecasts | Tests whether each revision should be accepted | Medium |
| Reports average forecast accuracy | Measures when revisions specifically cause harm | Low-medium |
| Uses uncertainty as model output | Uses uncertainty to choose whether to act | Medium |
7. Two-Week Pilot
- Week 1: Reproduce a baseline and build controlled relevant, irrelevant, and contradictory context cases.
- Week 2: Test simple gates and the proposed method. Produce a cost-versus-quality plot and a go/no-go decision.
- Stop condition: No improvement over the best simple baseline on both datasets.
Complete Examples
- LLM agents for time-series forecasting
- MLLM reasoning for text-to-image generation
- AI for scientific discovery
- Medical AI and surgical video
Evaluation
The scorecard checks evidence quality, opportunity clarity, novelty, plain-language explanation, reviewer criticism, experiment feasibility, and proposal readiness.
Limitations
- Prompt-only: It provides research instructions, not an automated research platform.
- No persistent memory: The temporary evidence map lasts only for the current context.
- No database-backed graph: It does not implement a knowledge graph or retrieval backend.
- Search-dependent: Poor or incomplete sources lead to poor opportunity maps.
- No novelty guarantee: A careful search reduces novelty risk but cannot prove that no similar work exists.
- Not a substitute for domain expertise: High-stakes fields still require expert review.
- Broad topics need narrowing: A useful answer may require focusing on one task, dataset, mechanism, or failure mode.
Repository Structure
.
|-- README.md
|-- SKILL.md
|-- examples/
| |-- 01-llm-agents-example.md
| |-- 02-mllm-t2i-example.md
| |-- 03-ai-for-science-example.md
| `-- 04-medical-ai-example.md
|-- eval/
| |-- test-prompts.md
| `-- output-scorecard.md
`-- LICENSE