research-orchestrator

name	research-orchestrator
description	Plan and execute deep research tasks with explicit scoping, phased search, source-tiering, multi-agent decomposition, red-team verification, and decision-ready synthesis. Use when the user asks for detailed research, broad/deep investigation, multi-round search, competitor/landscape analysis, technical deep dives, or wants a plan before execution and may benefit from subagents.

Research Orchestrator

Run research as a full deep-research workflow, not a one-shot answer.

This skill is Full Mode only. Do not silently downgrade it into a light search, a single-thread summary, or a quick memo. If the user explicitly invokes research-orchestrator, the default expectation is:

• longer-horizon research
• broader source coverage
• explicit planning before execution
• multi-agent / multi-lane decomposition
• an independent red-team pass
• structured synthesis with evidence separation

If the task is too small for this workflow, say so explicitly and ask whether the user wants a lighter workflow or a different skill. Do not silently execute a Lite or Standard version inside this skill.

Research orientation

This skill has only one execution mode: Full Deep Research.

You may still choose an orientation for the final output based on downstream use:

• exploration-oriented: map a space, build a taxonomy, identify open questions
• decision-support-oriented: support a recommendation or choice with tradeoffs and evidence
• executive-brief-oriented: produce conclusion-first, leader-friendly synthesis
• technical-due-diligence-oriented: stress-test technical claims, architecture, benchmarks, and risks

These are output lenses, not execution modes. They do not reduce the depth requirement. If helpful, read references/output-orientations.md.

Execution contract

If this skill is invoked, the minimum contract is:

1. Light search first
2. Execution plan before full run
3. Subagents or equivalent role-based lanes are required
4. A dedicated red-team lane is required
5. Source tiers must be handled explicitly
6. Final output must separate facts, inferences, and judgments
7. Final output must include a short execution header

If any part of this contract cannot be fulfilled because of tool/runtime constraints, tell the user during planning. Do not silently degrade.

Planning step is mandatory

Before doing heavy research, always present a compact planning block that covers:

• task understanding
• research orientation and why it fits
• target output and audience
• key subquestions
• scope boundaries
• source-tier plan
• subagent / lane topology and responsibilities
• red-team plan
• expected deliverable structure
• any known constraints or downgrade risks

If the topic is ambiguous, ask a small number of high-value clarification questions. If the topic is clear enough, proceed after the plan.

Light search before full run

Before locking the plan, do a quick search pass to identify:

• official / primary sources
• canonical names, aliases, dates, and entities
• likely comparison targets
• obvious controversies or moving parts
• whether the problem should be split differently than the default topology

Use this pass to refine the plan. Do not jump straight to final synthesis.

Subagents are mandatory in Full Mode

For this skill, subagents are not optional in normal operation. At minimum, use role-based decomposition with clear differentiation. Do not spawn multiple agents that search the same angle with no role distinction.

Minimum required topology

Use this baseline unless there is a strong reason to adapt it:

1. Planner
   • define the frame
   • split the work into dimensions
   • set success criteria and unknowns
2. Primary-source scout
   • collect Tier 1 sources
   • normalize names, dates, objects, and primary evidence
3. Domain lane A
   • first major dimension
4. Domain lane B
   • second major dimension
5. Red-team / skeptic
   • search for contradictions, adverse evidence, weak evidence, outdated claims, and overreach
6. Main synthesizer
   • merge, deduplicate, classify evidence, and write the final view

You may add more lanes when the topic justifies it, but do not go below this topology without explicitly telling the user why.

When to expand beyond the minimum topology

Add more lanes when:

• the topic naturally splits into 3+ distinct dimensions
• breadth matters as much as depth
• multiple expert angles are needed
• verification is especially important
• the user explicitly asks for very broad, deep, or multi-angle research

If helpful, read references/multi-agent-topologies.md.

Red-team lane is mandatory

Every serious research memo produced via this skill must include an independent red-team lane. This lane should actively look for:

• contradictory claims
• missing baselines
• outdated information
• marketing inflation / hype narratives
• adverse evidence
• weak or circular sourcing
• explanations that sound plausible but do not match the timing or evidence

Do not reduce the red-team pass to a token paragraph. It should materially pressure-test the main thesis.

Minimum red-team deliverable standard

At minimum, the red-team output should cover:

1. Evidence gaps
   • which important claims are still weakly supported, indirectly supported, or not yet well evidenced
2. Alternative explanations
   • what competing interpretation could also explain the observed facts
3. Boundary conditions
   • where the main conclusion may fail, narrow, or stop generalizing

If no strong counterexample is found, do not just say "no issues found". Explicitly state:

• which challenge directions were checked
• which strong rebuttals were not found
• which uncertainties still remain unresolved

Source tiers are required

Distinguish source quality in both research and output:

• Tier 1: official docs, papers, repos, benchmarks, launch posts, filings, direct datasets, primary disclosures
• Tier 2: credible analysis, industry coverage, high-quality technical writeups, expert commentary
• Tier 3: community discussion, social posts, anecdotal operator feedback, forum sentiment

Prefer Tier 1 for factual claims. Use Tier 3 mainly for hypothesis generation, sentiment, and edge-case signals. Do not let Tier 3 drive major conclusions unless clearly labeled and corroborated.

Evidence model is required

In the final synthesis, separate:

• facts: directly supported by sources
• inferences: reasoned conclusions built from facts
• judgments: strategic views, recommendations, or prioritization

Also include confidence levels where useful. Do not collapse facts, inferences, and opinions into one bucket.

Default split patterns

Use these as defaults when planning lane structure.

Market / macro / investing topic

• market facts and timeline
• driver analysis and transmission mechanisms
• sector / style / asset-chain implications
• red-team

Technical topic

• architecture / implementation
• benchmarks / eval quality
• ecosystem / tooling / adoption
• red-team

Product / company topic

• product / capability map
• market / GTM / adoption
• competition / positioning
• red-team

Landscape topic

• taxonomy / segmentation
• top players and positioning
• evidence by segment
• strategic implications
• red-team

Output requirements

Prefer structured, decision-useful outputs over note dumps. The final output should usually contain:

• short execution header
• executive summary
• key questions
• key findings
• source-tier-aware evidence
• disagreements / uncertainties
• facts vs inferences vs judgments
• implications / recommendations
• next-step research suggestions

Required execution header

At the top of the final output, include a compact execution header covering:

• Mode: Full Deep Research
• Orientation: exploration / decision-support / executive-brief / technical-due-diligence
• Topology: planner + primary-source + domain lanes + red-team + synthesis
• Subagents used: yes/no and how many
• Red-team pass: completed / constrained
• Source mix: Tier 1 dominant / mixed / etc.
• Evidence model: facts / inferences / judgments separated

Failure modes to avoid

Do not do any of the following:

• silently downgrade to a lighter workflow
• write a long report while skipping subagents and still treat it as full deep research
• skip the primary-source lane
• skip the red-team lane
• spawn multiple agents with no role distinction
• summarize material without making judgments where the task requires them
• mix facts, inferences, and opinions into a single undifferentiated narrative

Prompt pattern

Use this shape when formulating or refining a deep-research request:

• subject
• objective
• audience
• output format
• constraints / scope boundaries
• source preferences
• required subagent topology
• required verification / skepticism level
• expected deliverable structure

If helpful, read references/prompt-patterns.md for reusable request templates.