Español

Don't Ask AI to Research. Ask It Five Questions

Stop asking AI to research. Ask it five questions.

A single "research this topic" prompt gives you a single shallow pass: the model skims the surface, picks a few generic points, and returns something that reads fine but has no depth and no sources you can check. Depth doesn't come from one big ask. It comes from structure, decomposing the topic, searching each part against real sources, and stitching the findings back together.

Why one big ask stays shallow

A broad prompt forces the model to cover an entire topic in one context window and one reasoning pass. It has to be general to be complete, so it trades depth for breadth and rarely digs into any single thread. It also has no reason to cite: nothing in "research X" demands that each claim carry a source, so you get assertions you have to re-verify yourself. The fix is to change the shape of the work, not the size of the prompt.

The mechanism: decompose, fan out, synthesize

Turn one vague ask into a small research pipeline.

1. Decompose. Break the topic into three to five specific sub-questions. Each should be answerable on its own and narrow enough to go deep.
2. Fan out. Dispatch each sub-question to its own subagent, running in parallel. Every subagent searches the web and reads the actual sources for its slice, not the whole topic.
3. Synthesize. Collect the sub-answers and merge them into one report where every claim carries the source it came from.

Topic: "Is X production-ready for our use case?"

  Q1: maturity + release cadence?   ─┐
  Q2: known failure modes?           ├─ parallel subagents ─→ synthesize
  Q3: performance vs alternatives?   │    (search + read)      one cited
  Q4: licensing + cost?              │                          report
  Q5: who runs it in production?    ─┘

One big ask vs fan-out

The payoff is broader coverage, every claim traceable to a source, and no single reasoning thread becoming the bottleneck.

Power moves

• Write sub-questions that don't overlap. Orthogonal questions maximize coverage; near-duplicates waste a subagent.
• Make each subagent cite as it reads. Require a source per finding at the subagent level, so the final synthesis is cited by construction, not patched afterward.
• Run an adversarial pass in synthesis. Have the synthesizer flag claims where subagents disagree or a source is weak, rather than smoothing them over.
• Keep sub-questions concrete. "Known failure modes of X in production" goes deeper than "tell me about X."

Resources

• Anthropic — How we built our multi-agent research system
• Anthropic — Building effective agents
• OpenAI — Orchestrating agents and tools
• Lewis et al. — Retrieval-Augmented Generation