Research Review Skill Factory

Use this meta-skill to build a custom review skill for a specific research area, problem family, or method combination. It is broader than a manuscript-specific builder: the generated child skill should help review future papers in the selected area.

Core Idea

Create a field/problem-specific reviewer skill:

research area + problem set -> area profile -> OpenReview queries -> reviewer concern patterns -> custom area reviewer skill

Examples:

• ssfl-diffusion-representation-reviewer-openreview
• federated-ssl-privacy-reviewer-openreview
• spectral-representation-theory-reviewer-openreview
• llm-agent-benchmark-reviewer-openreview

Workflow

1. Define the research area and problem set

   • Ask for or infer the area scope: narrow field, parent fields, problem family, method families, theory objects, experiment settings, and target venues.
   • Use references/research_area_profile_schema.md.
   • Preserve narrow terms before broad terms.

2. Generate OpenReview query plan

   • Create 8-20 queries covering the exact area phrase, subproblems, method families, theory or benchmark keywords, closest baseline families, and broader fallback fields.
   • Check the current date and select the current ICLR year plus two previous public ICLR years unless the user specifies years.

3. Retrieve public OpenReview evidence

   • Use:

python scripts/fetch_openreview_field_evidence.py --field "<query>" --years <Y1> <Y2> <Y3> --output "<evidence-dir>/<query-slug>"

• Collect reviewer concerns from accepted, rejected, withdrawn, and desk-rejected public submissions when available.
• Use author responses only from accepted papers by default.

4. Synthesize an area review-response bank

   • Cluster reviewer concerns by category.
   • For each pattern, record trigger terms, reviewer concern, accepted-paper response pattern, what future papers in this area must show, and representative evidence.
   • Keep direct quotes short; paraphrase patterns and cite forum URLs.

5. Generate the child area reviewer skill

   • Use scripts/init_research_area_review_skill.py with a filled area profile JSON.
   • The generated child skill must include SKILL.md, agents/openai.yaml, references/research_area_profile.md, references/openreview_review_response_bank.md, references/review_output_contract.md, references/subtle_logic_flaws.md, LICENSE.txt, and _meta.json.

6. Validate and package

   • Run quick_validate.py on the child skill.
   • Run syntax checks on scripts.
   • Package the child skill only after confirming there are no raw evidence caches, PDFs, manuscripts, pycache, or private data.

Generated Child Skill Requirements

The child skill must instruct future reviewers to:

• classify a submitted paper inside the target research area;
• retrieve the local area review-response bank before writing review comments;
• generate area-specific reviewer concerns and rebuttal/revision guidance;
• cite OpenReview precedent with year, status, title, forum URL, and note type;
• audit novelty, soundness, baselines, reproducibility, A+B incrementality, and subtle logic flaws;
• provide light, moderate, and major revision paths.

Evidence Rules

• Never fabricate OpenReview titles, forum IDs, decisions, scores, or author responses.
• Treat OpenReview evidence as precedent, not as law.
• Do not include raw review dumps in the generated child skill.
• If evidence is sparse, label the bank as limited evidence and include a broader fallback area.

References

• references/research_area_profile_schema.md: area/problem profile schema.
• references/openreview_area_evidence_workflow.md: retrieval and synthesis protocol.
• references/generated_area_review_skill_contract.md: generated child skill contract.
• references/subtle_logic_flaws.md: reusable hidden-weakness checklist.