Research Ideation

• topic or problem statement
• target task or scenario
• hard constraints and non-goals
• what would count as a better idea
• known pain points or hypotheses
• optional target venue, dataset, metric, or engineering preference

2. Local papers

Use papers/**/*.md as the default durable paper source for this skill.

Rules:

• read only the subset relevant to the user brief
• do not invent details missing from the notes
• if the local notes are weak, explicitly downgrade confidence

3. Optional online paper search

Use online paper search when one or more of the following are true:

• the user explicitly asks to search for papers online
• the local paper set is too small to bound novelty risk
• the local notes appear outdated relative to a fast-moving topic
• the local notes mention adjacent work but not the closest current neighbors

Rules:

• prefer Exa MCP for paper discovery
• use online search to widen the candidate set, not to fabricate deep reading
• label whether a claim comes from local notes, online abstracts, or fuller paper evidence

Workflow

1. Read the user brief first.

   • Extract problem, constraints, assumptions, non-goals, and what would count as a better idea.

2. Clarify blocking gaps before reading papers when needed.

   • If the brief is missing any of these fields in a way that would change the ideation result, use request_user_input before continuing:
     • target task or scenario
     • hard constraints
     • non-goals
     • what would count as a better idea
   • Ask one batched clarification round only.
   • Ask only the 1 to 3 highest-leverage questions.
   • Treat the user's answers as an in-session supplement to the brief.
   • Do not require the user to rewrite the Markdown file before continuing.

3. Build a local paper map.

   • Find the relevant files in papers/**/*.md.
   • Extract method family, core mechanism, assumptions, strengths, limitations, and reusable patterns.

4. Expand the paper map online when needed.

   • Search for recent or adjacent papers with Exa MCP.
   • Use titles, abstracts, venues, and dates to identify the strongest nearest neighbors.
   • Keep evidence tiered: online discovery is not the same as full-paper verification.

5. Compare the brief against the grounded paper map.

   • Identify what is already covered.
   • Identify what is missing, contradictory, weakly evaluated, or overly narrow.

6. Produce a grounded gap view.

   • Distinguish real capability gaps from narrative or packaging gaps.
   • Call out where the evidence is strong and where it is incomplete.
   • Separate local-note evidence from online-search evidence.

7. Form one or more testable candidate questions.

   • Keep the question tied to a concrete task, comparison target, metric, and failure condition.

8. Draft a minimal validation plan.

   • State what evidence would justify implementation.
   • State what would falsify the idea early.

Agent Use

• Use paper-miner when the local paper set is large and heterogeneous.
• Use literature-reviewer when the job is to synthesize and contrast many local paper notes.
• Use online paper search when the user asks for broader coverage or when the local corpus is too weak to bound the problem responsibly.

Source Rules

• Base every core judgment on the user brief, local paper notes, and online paper search results when used.
• Treat local paper notes as the strongest reusable evidence layer inside this repo.
• Treat online paper search as a widening layer unless the paper has been read deeply enough to support stronger claims.
• If a claim cannot be supported by the local notes, say so directly.
• If a claim is supported only by online search snippets or abstracts, say so directly.
• Do not turn missing evidence into a confident novelty claim.

Output Rules

• Respond in Chinese.
• Keep the answer summary-first.
• Make reused patterns and changed assumptions explicit.
• Make clarification-derived assumptions explicit when request_user_input was used.
• State uncertainty and missing evidence explicitly.
• Label the evidence source when both local notes and online paper search are used.
• Prefer outputs such as gap lists, candidate questions, risk tables, and minimal validation plans.

Failure Handling

• Missing user brief: ask for a valid Markdown path.
• Materially underspecified brief: use one batched request_user_input round before paper mapping.
• Clarification still unavailable: stop and state which blocking fields are still missing.
• Empty papers/: say that local paper grounding is unavailable and switch to online paper search with downgraded confidence, or stop if the user wants local-only grounding.
• Irrelevant local notes: say that the current corpus does not support the requested direction well.
• Online search unavailable: continue with local notes only and clearly mark the literature boundary as incomplete.
• Weak novelty evidence: avoid claims such as first, novel, or state of the art.