By name: research-brainstorm description: Brainstorm follow-up research ideas from papers. Use when asked to generate research directions or ideas.
Research Brainstorming
When to use
- User says "what could we do next based on this paper?"
- User asks for research directions, follow-ups, or extensions
- User wants to identify gaps in existing work
Workflow
Understand the paper deeply before brainstorming
- What's the core contribution? What's actually new?
- What assumptions does it rely on?
- What did they NOT do, and why?
- What are the stated limitations?
Identify gaps from multiple angles:
- Methodology: can the approach be applied to different domains/modalities?
- Scale: does it work at 10x/100x the tested scale?
- Assumptions: what happens when key assumptions are relaxed?
- Combination: can this be combined with orthogonal recent work?
- Failure modes: where does the method break down?
Generate 3-5 concrete research directions, each with:
- One-line description: specific and actionable
- Gap: what's missing in the current work that this addresses
- Approach sketch: 2-3 sentences on how you'd actually do it
- Feasibility: easy (weeks, existing tools) / moderate (months, some new infrastructure) / hard (6+ months, fundamental challenges)
- Resources needed: data, compute, specific expertise
Quality bar: every idea must be specific enough to start working on
- Bad: "improve the model's performance"
- Bad: "apply to other domains"
- Good: "replace the cross-attention in the fusion module with a mixture-of-experts gate, since the current design scales quadratically with the number of modalities — test on the AudioSet + ImageNet joint benchmark"
Prioritize by the intersection of novelty and feasibility — the best ideas are those that are clearly doable and clearly haven't been done
Formatting for Slack
- Number the ideas
- Bold the one-line description
- Keep feasibility assessments honest — don't oversell
- If an idea requires specific expertise the lab may not have, say so