Research Scoping

Guidelines for turning research ideas into well-scoped task trees. Apply at root (task "0") and when refining broad sub-tasks.

Scoping Principles

1. Concrete over vague: "Compare Python vs JS for backend" not "Research languages"
2. Deliverable-focused: Each task should produce a clear artifact (document, list, config, report)
3. Depth-appropriate: Root level = high-level phases (2–6); deeper = specific sub-steps
4. Dependency-aware: Order by data flow. Gather before analyze. Setup before execute.
5. Feasible scope: Avoid "complete the entire research" as a single task; break into phases.
6. Refine ambiguous ideas: If the idea is "调研某技术" (research some tech), clarify: which tech? what aspects? what output?

Common Research Structures

Domain	Typical Phases	Notes
Literature review	Search → Filter → Synthesize → Report	Each phase has distinct methodology
Technical comparison	Scope → Research A → Research B → Compare	Parallel research, then synthesis
Experiment	Hypothesis → Setup → Run → Analyze	Sequential, data flows forward
Survey/Interview	Design → Conduct → Analyze → Report	Design defines Conduct; Conduct feeds Analyze
Benchmark/Evaluation	Define metrics → Setup → Run → Aggregate	Similar to Experiment
Documentation	Outline → Draft → Review → Finalize	Iterative refinement
Gap analysis	Define scope → Gather current state → Identify gaps → Recommend	Feasibility, maturity assessment

Granularity Guidelines

• Root (depth 0): 2–6 phases. Each phase = distinct methodology or major deliverable.
• Mid-level: 2–4 sub-tasks per phase. Sub-tasks = concrete steps within that methodology.
• Leaf (atomic): Single focused session, one clear output. No further decomposition.

Refinement Tactics

When the idea is too broad:

• Narrow by domain: "AI in healthcare" → "AI for medical image diagnosis"
• Narrow by output: "Research X" → "Produce comparison report on X vs Y"
• Add constraints: "Survey frameworks" → "Survey Python web frameworks (2020+)"
• Split by dimension: "Evaluate tools" → "Evaluate by performance" + "Evaluate by ecosystem"

Atomicity Hints

• Atomic: Single search, single analysis, single write-up, single decision, one clear deliverable
• Non-atomic: Multiple methodologies, distinct deliverables, handoff points, "and then" in description

Paper Agent Boundary

Do not create a task for the final research paper. The Paper Agent generates the paper from task outputs after execution. Tasks may produce summaries, synthesis reports, comparison outputs, or literature reviews—these are appropriate. The Paper Agent composes them into the final paper.

Red Flags

• Task description contains "and" linking two distinct activities → likely non-atomic
• "First X, then Y" → two phases, decompose
• No clear deliverable → refine description or decompose
• "调研" (research) without scope → add scope or decompose into scope + gather + synthesize
• Vague "分析" (analyze) without input spec → clarify what is being analyzed and from where