devjarus/deep-research

Deep Research

How to run research that holds up: parallel breadth, then adversarial depth. The orchestrator is the lead agent; this skill is the method it (and its research subagents) follow. See ${CLAUDE_PLUGIN_ROOT}/protocols/research.md for the actor choreography.

The shape

decompose → fan out (parallel) → interleaved-think per result → verify (refute) → synthesize (cited)

Single-context, sequential research is the failure mode: it's slow, path-dependent (each query biased by the last), and rarely refutes itself.

1. Decompose (think hard first)

Engage extended thinking before any search. Break the question into 2–5 independent sub-questions — ones that don't need each other's answers. If two sub-questions must be answered in order, they're one investigation, not two.

Good decomposition for "best queue for our Node service":

• throughput/latency profile of candidates at our scale (Architecture)
• ops/cost of managed vs self-hosted (Deployment/Cost)
• delivery guarantees + dedupe semantics (Data)

2. Fan out in parallel

One investigator per sub-question, dispatched together. Isolated contexts mean independent exploration trajectories — less path dependency, broader coverage. Each investigator calls tools in parallel where it can.

Tool per job:

• Glob/Grep — what the codebase already does
• Context7 — current library/framework/API docs (memory is stale)
• Exa — neural web search for comparisons, benchmarks, postmortems
• WebSearch/WebFetch — current events, vendor docs, pricing

3. Interleaved thinking on every result

Don't fire the next query blind. After each tool result, reason: Did this answer the sub-question? What does it contradict? What's the next sharpest query? Refine. A good investigator's 4th query is shaped by its first three results.

4. Verify adversarially

The most important step, and the one usually skipped. For every load-bearing claim, try to break it:

• Find a second independent source.
• Look for a counter-example or a "but not when…".
• Check recency — is this true for the current version?

Tag each surviving claim high / medium / low confidence. Record claims you refuted — they stop the answer from drifting back to a plausible-but-wrong default.

5. Synthesize, cited

Reconcile across investigators. Resolve contradictions explicitly — name the winner and why, don't average. Every recommendation carries its source. Surface what you couldn't resolve as open questions; a known gap beats false confidence.

Anti-patterns

• One agent doing 15 sequential searches in its own context (slow, biased, no refutation).
• Accepting the first source as truth.
• A claim with no source.
• Synthesizing before verifying.
• Decomposing a 1-fact question into a 5-agent fan-out (match effort to the question).