cross-paper-skill-synthesizer

name	cross-paper-skill-synthesizer
description	Cluster, deduplicate, generalize, and rank proof-pattern SkillCandidates or SkillCards extracted from multiple papers. Use when a coding agent should synthesize single-paper extractions into a domain method map, merged SkillCards, and a synthesis report for an AI4Math skill library.

Cross Paper Skill Synthesizer

Purpose

Use this skill after multiple papers have been processed by paper-to-skill-extractor. The goal is to build a domain method map and library-ready SkillCards, not to summarize papers.

Inputs

Accept any mix of:

• extractions/<paper_id>/skill_candidates.json
• extractions/<paper_id>/skill_cards/*.yaml
• Paper metadata files.
• User notes about which skills should be merged or rejected.
• Optional human_feedback_state.json from the research project.

Workflow

1. Load all SkillCandidates and SkillCards.
2. If human_feedback_state.json exists, read focus_updates, negative_preferences, skill_decisions, and next_step_directives before clustering.
3. Preserve source evidence from every paper. Do not drop line references.
4. Cluster candidates by proof pattern, intent, assumptions, inputs, outputs, and core steps.
5. Deduplicate near-identical skills while keeping all source references.
6. Generalize paper-specific notation into reusable method language.
7. Identify conflicts: incompatible assumptions, different conclusions, unsafe generalizations, or missing evidence.
8. Rank clusters for library readiness.
9. Write domain_method_map.yaml, merged SkillCards, and synthesis_report.md.
10. In synthesis_report.md, explain which user feedback affected merge, reject, or library-readiness decisions.

Required References

Read only as needed:

• references/synthesis_rules.md: clustering, merging, and safety rules.
• references/schemas.md: required output fields.

Use the templates:

• assets/domain_method_map.template.yaml
• assets/merged_skill_card.template.yaml

Output Files

outputs/<research_project_id>/synthesized_skills/
├── domain_method_map.yaml
├── merged_skill_cards/
│   └── <skill_name>.yaml
└── synthesis_report.md

Synthesis Rules

• Never remove source evidence when merging.
• Do not generalize beyond the shared assumptions unless marked needs_human_review.
• Separate ready_for_library, needs_human_review, too_specific, duplicate, and unsafe_generalization.
• Keep rejected or unsafe clusters in the report so future reviewers understand the decision.
• Prefer fewer, stronger SkillCards over many thin variants.
• Apply user skill_decisions when grouping, merging, rejecting, or marking skills for revision.
• Use human focus updates to prioritize the domain method map, but keep conflicting or unsafe clusters visible instead of silently dropping them.

Completion Check

Before finishing, confirm:

• Every merged SkillCard has source references to original paper.md files.
• domain_method_map.yaml records clusters, representative skills, and statuses.
• The report explains conflicts, risks, and human-review items.