| name | research-idea-generation |
| description | Generate and score candidate LLM research ideas from sparse user goals. Use when a project starts with no concrete benchmark, dataset, or training recipe and the agent needs to propose plausible, testable directions. |
| version | 1.0.0 |
| author | Hermes Research Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["research","ideation","literature","zero-spec"]}} |
Research Idea Generation
Use this skill when the user gives a vague or open-ended research goal.
Workflow
- Search recent literature and identify recurring limitations, evaluation gaps, or underexplored settings.
- Produce 3 candidate directions.
- For each direction, write:
- one-sentence thesis
- why it might work
- the likely benchmark or eval target
- expected cost/risk
- Score each direction on:
- novelty
- feasibility
- cost
- evaluation clarity
- Choose one and explain the decision in terms of upside versus execution risk.
Constraints
- Prefer ideas that can be tested with a clear baseline and measurable success criterion.
- Avoid open-ended ideas that require undefined new infrastructure for v1.
- Record all candidates and the chosen direction with
research_state.
