| name | research-orchestrator-v2 |
| description | Plan and execute deep research with explicit scoping, light-search-first planning, source-tiering, role-based parallel decomposition, red-team verification, and decision-ready synthesis. Use when the user asks for deep research, broad investigation, multi-round search, landscape mapping, competitor analysis, technical diligence, executive briefs, recommendation support, plan-first research, or work that should benefit from subagents instead of a one-shot answer. |
Research Orchestrator V2
Run research as a staged system, not a single long answer. Optimize for decision usefulness, evidence quality, contradiction-finding, and reusable outputs.
1. Frame before searching hard
Before heavy research, lock five things:
- the real decision or downstream use
- target audience and output shape
- in-scope vs out-of-scope boundaries
- key unknowns / subquestions
- whether this is narrow enough for one thread or broad enough for parallel lanes
If the request is underspecified, ask a small number of high-leverage questions. If it is clear enough, proceed.
2. Do a light search before committing to a plan
Never produce a blind research plan.
Use a quick first pass to identify:
- canonical names, aliases, and keywords
- likely Tier 1 sources
- obvious comparison targets
- moving parts, controversies, or timeline issues
- whether the topic is broad enough to justify subagents
Then present a short plan unless the user clearly asked to skip straight to execution.
If you need mode guidance or a planning template, read references/research-modes.md and references/research-plan-template.md.
3. Choose the right research mode
Use one dominant mode:
- exploration: map the space before judging
- decision-support: help choose, compare, or recommend
- executive-brief: compress into a leader-ready brief
- technical-due-diligence: test technical substance, benchmarks, and risk
Infer the mode from downstream use, not surface wording.
4. Split by role, not by arbitrary parallelism
When the task is broad, high-stakes, or explicitly deep, use role-based decomposition.
Default lanes:
- planner — frame, split, and define success criteria
- primary-source scout — collect Tier 1 evidence and canonical artifacts
- domain lane(s) — investigate distinct dimensions such as architecture, product, market, adoption, GTM, or ecosystem
- skeptic / red-team — look for contradictions, missing baselines, adverse evidence, or stale claims
- synthesizer — merge, judge, classify uncertainty, and produce the final deliverable
Do not spawn multiple agents that all search the same angle with slightly different wording.
If needed, read references/multi-agent-topologies.md.
5. Use source tiers deliberately
Classify evidence while researching and again while writing:
- Tier 1: official docs, papers, repos, benchmarks, filings, launch posts, source code, vendor docs
- Tier 2: credible technical blogs, industry analysis, strong media coverage, expert writeups
- Tier 3: community discussion, social posts, anecdotal operator reports
Rules:
- anchor factual claims in Tier 1 whenever possible
- use Tier 2 to contextualize and compare
- use Tier 3 for sentiment, practitioner signal, and hypothesis generation
- do not let Tier 3 dominate a serious memo unless the topic itself is community sentiment
If needed, read references/evidence-and-output-rules.md.
6. Require a red-team pass for serious work
For any nontrivial memo, explicitly search for:
- contradictory evidence
- benchmark or demo inflation
- outdated claims
- missing baselines
- deployment friction
- weak adoption evidence
- reasons the obvious conclusion might be wrong
If evidence conflicts, preserve the disagreement. Do not force fake certainty.
7. Separate facts, inferences, and judgments
In the final synthesis, distinguish:
- facts — directly supported by sources
- inferences — conclusions derived from evidence
- judgments — recommendations, bets, or strategic interpretation
Add confidence levels when useful.
8. Tailor the final output to the real use
Prefer decision-ready structure over raw notes.
Common sections:
- executive summary
- task framing and scope
- key findings
- comparison matrix or dimension table
- disagreements / uncertainty
- confidence levels
- implications / recommendations
- next-step research lanes
Mode emphasis:
- exploration -> taxonomy, landscape map, open questions, follow-up lanes
- decision-support -> recommendation, tradeoffs, risks, action options
- executive-brief -> 3-7 key judgments, concise evidence bullets, next actions
- technical-due-diligence -> architecture, evidence quality, benchmark scrutiny, technical risks, unknowns
If needed, read references/output-shapes.md.
9. When not to use this skill
Do not use this skill for:
- a narrow factual lookup answerable from 1-3 sources
- summarizing one article or one PDF without broader research intent
- lightweight brainstorming with no evidence requirement
- simple web search where no staged workflow is needed
10. Prompt shaping pattern
When reformulating the task, try to make these fields explicit:
- subject
- objective
- audience
- output format
- scope constraints
- source preferences
- whether subagents are warranted
- required skepticism / verification level
If useful, read references/research-plan-template.md for a reusable plan shape.
