Research Survey And Idea

1. Build a topic profile from the user question.
2. Retrieve a high-recall candidate paper pool.
3. Screen into core and extended paper sets.
4. Download and parse core papers.
5. Build structured paper cards with evidence anchors where possible.
6. Synthesize technical routes and evolution paths.
7. Generate multiple ideas and challenge them before recommendation.
8. Select one top idea.
9. Write a proposal-style paper draft for the top idea.

Read workflow.md before running the full process.

Use Scripts

Use scripts for deterministic or repetitive steps:

• initialize run directories
• retrieve and merge paper metadata
• deduplicate and screen results
• fetch PDFs
• parse papers
• assemble deliverables from artifacts

Do not turn these steps into long free-form prompt chains if a script exists.

Use Prompts

Use prompt files for high-judgment steps:

• topic-expansion.md
• paper-analysis.md
• route-synthesis.md
• idea-generation.md
• draft-writer.md

Load References As Needed

Always read:

• workflow.md
• outputs.md

Read when relevant:

• gotchas.md before synthesis, ideation, or drafting
• ai-llm.md for AI / LLM topics
• computer-use-agents.md for Computer-Use Agent topics

Keep SKILL.md lean. Load detailed references only when the current step needs them.

Outputs

Human-readable deliverables belong in deliverables/.

Machine-readable artifacts belong in artifacts/.

The exact run-directory contract is documented in outputs.md.

Preserve artifacts so research-verification can audit coverage, evidence support, novelty risk, and draft integrity. If the user asks for trust checking, or if confidence is visibly low, invoke research-verification.

Constraints

• Do not present generic background as if it were evidence-backed synthesis.
• Do not let unsupported claims enter the final survey.
• Do not treat weak sources as anchors for core technical claims in v1.
• Do not recommend an idea without a challenge pass.
• Do not fabricate experiments, citations, or empirical results.
• Write the draft as a proposal-style or discussion-style paper if no real experiments exist.
• Prefer concise, traceable outputs over verbose but ungrounded narration.