| name | innovation-paper-finder |
| description | Search for papers that may provide new methods, new theory, sharper bounds, different assumptions, proof techniques, lower bounds, abstractions, or cross-domain transfers for a math research direction. Use when a coding agent should look beyond nearest-neighbor related work and generate innovation_candidates.json with reasons and risks. |
Innovation Paper Finder
Purpose
Use this skill to find papers that may create new research angles. This is not ordinary related-work retrieval. Search for transferable ideas, proof techniques, sharper assumptions, negative results, and adjacent-domain methods.
Inputs
outputs/<research_project_id>/research_profile.json
- Optional
outputs/<research_project_id>/human_feedback_state.json
- Optional
candidate_papers.json, seed papers, or user constraints.
Workflow
- Read the research profile and any existing candidate list.
- If
human_feedback_state.json exists, read it before selecting innovation routes. Apply focus_updates, negative_preferences, and next_step_directives.
- Identify innovation routes: new proof technique, sharper bound, weaker assumption, cross-domain transfer, lower bound, alternative formulation, or new abstraction.
- Use the active agent's available search or browsing capability for each route.
- Prefer primary paper pages and open full text where available.
- For each candidate, explain why it may help and what research angle it suggests.
- Record risks such as weak relevance, high reading cost, speculative transfer, or paywalled access.
- Write
innovation_candidates.json using references/schemas.md.
- Write
innovation_report.md with recommended next actions, applied human feedback, and checkpoint questions.
Required References
Read only as needed:
references/innovation_signals.md: signal taxonomy and tests.
references/schemas.md: required output fields.
Use the template:
assets/innovation_candidates.template.json
Output Files
outputs/<research_project_id>/
├── innovation_candidates.json
└── innovation_report.md
Search Rules
- Deliberately search adjacent fields, not only exact keyword neighbors.
- Include at least one route targeting assumptions and one route targeting proof techniques when possible.
- Prefer routes that match explicit user focus updates, such as proof technique, coupling, error bounds, or lower-bound directions.
- Down-rank speculative leads that match user negative preferences unless they are included as a clearly labeled contrast case.
- Keep speculative ideas, but label them as speculative.
- Do not download full text unless the user selected the paper or explicitly requested download.
- Do not bypass paywalls.
Completion Check
Before finishing, confirm:
innovation_candidates.json is valid JSON.
- Every candidate has an innovation signal, possible research angle, risk note, and recommended action.
- The report separates strong leads from speculative leads.