Research Orchestration

Research Orchestration

Use When

• Use as the first step on any non-trivial research project, planned research wave, multi-cohort investigation, gap-fill pass, verification pass, or synthesis pass.

Do Not Use When

• Do not use for a single known fact lookup, URL check, or short target read that does not need orchestration.

Required Inputs

Workflow

1. Run the type, discipline, and reading-mode routers before wave dispatch.
2. Select companion skills for evidence discipline, research design, methods, source evaluation, and output form.
3. Dispatch planned waves, verify before merging, and reserve final synthesis for the orchestrator.
4. Stop when scope, evidence rules, or verification ownership is unresolved.
5. Recover from unavailable agents or providers by running sequentially and preserving unassessed gaps.

Quality Standards

• Research waves must be scoped, source-aware, evidence-disciplined, and designed to produce insight rather than volume.

Anti-Patterns

• Do not run a single mega-search when cohorts, source classes, or verification needs differ. Fix: split non-overlapping waves.
• Do not delegate cross-cohort synthesis to wave agents. Fix: reserve synthesis for the orchestrator.
• Do not merge an unverified wave. Fix: quarantine claims and schedule verification.
• Do not hide provider failure. Fix: label the gap and preserve the attempted path.
• Do not optimise speed at the expense of source diversity. Fix: keep a quality guardrail.

Outputs

References

• Use the routers and companion-skill table below.
• Use references/kaizen-research-loop.md after each wave and before release.

The first skill loaded on any research project. Decides what kind of research is being run, what discipline its language and sources come from, what reading mode each retrieved source warrants, and how to dispatch the work as planned waves.

Step 0 — pick the type, the discipline, the reading mode

Before any wave fires, run three routers (each lives in references/):

Skipping these is the most common cause of research that produces volume but not insight.

Companion-skill router (load in addition to this skill)

The wave model

A research engine never runs as a single search. It runs as planned waves:

1. Wave 1 — broad sweep. One agent per cohort, broad query, ~50 sources target.
2. Wave 2 — gap fill. Specific gaps each Wave 1 agent flagged; deeper academic + regulatory sources.
3. Wave 3 — verification. Cross-source triangulation; verbatim-quote extraction; URL liveness check.
4. Wave 4 — synthesis. Cross-cohort patterns; product/policy inferences.

This skill defines when each wave fires and how outputs are stored.

Kaizen after every wave

Every wave must leave a short learning record: observation, reproducible baseline, one hypothesis, smallest reversible experiment, guardrail, result, failed-path result, and standardisation decision. Useful measures include source admission rate, verified-claim coverage, unresolved-claim count, contradiction count, duplicate-search rate, and time-to-verified-finding. These are measures to collect, not values to invent. If the experiment improves speed but weakens source diversity, independence, or verification coverage, reject it and preserve the prior standard.

When to use

Trigger on:

• "Research X across [region]" — dispatch a Wave 1 sweep
• "Now do Y for [adjacent group]" — parallel Wave 1 on a new cohort
• "Do another thorough pass" — Wave 2 gap fill
• "Synthesise across cohorts" — Wave 4
• "Audit our research for what we missed" — Wave 3 + Wave 4

Do not use for: single-question lookups, fact-checks, or known-target reads.

Wave Choice Rules

• One cohort per agent. Don't ask one agent to research students AND landlords. Split.
• Run waves in parallel where independent. A student-side gap-fill and an owner-side gap-fill have no shared state — fire them simultaneously.
• Cap each agent at ~50 distinct sources. Past that, marginal value drops sharply; better to fire a Wave-2 agent.
• Never delegate synthesis to research agents. The orchestrator (you) does the synthesis. Research agents return raw findings, not conclusions.
• Always brief the agent on what's already known. Wave-2 agents must see Wave-1 themes so they don't re-confirm; they're paid to find new material.

Standard agent brief structure

Every research-agent prompt should contain:

1. Goal — one sentence
2. Scope — geography, cohort, time-window
3. Audience / decision context — who will use it and what decision it must support
4. Out-of-scope — what NOT to research (avoids duplicate work)
5. Themes to cover — explicit numbered list
6. Sources to mine — named outlets, repositories, platforms
7. Deliverable shape — markdown sections, source-count target, quote requirements
8. Hard constraints — exclusions (e.g., "do NOT cover topic X"), language, format

If the brief is missing any of those, the output will be uneven.

Orchestrator outputs

After each wave, the orchestrator writes:

• <cohort>/research/pain-points-report.md (or complaints-report.md for student-style cohorts)
• <cohort>/research/sources.md — annotated source list with gaps section
• <cohort>/research/quotes.md — verbatim quotes
• <cohort>/analysis/themes.md — sub-theme taxonomy
• <cohort>/analysis/by-country.md (or by-segment, depending on dimension)
• <cohort>/opportunities/product-ideas.md

Wave 2 outputs append a # Pass 2 — Gap-fill addendum section to the existing files rather than overwrite.

Wave Failure Modes

• Single-shot research with one mega-prompt — produces shallow coverage, missing perspective splits
• Re-running Wave 1 instead of running Wave 2 — wastes context, duplicates known
• Delegating cross-cohort synthesis — only the orchestrator has all four cohort outputs in context
• Forgetting hard constraints in Wave-2 briefs — agent re-introduces excluded scope

See also

• research-techniques — gap analysis, mini-analysis, crosswalk matrix, cross-cohort synthesis, reference interview, controlled vocabulary, search-operator grammar, pearl-growing, brachiation, etc. (load when a Wave needs a specific technique)
• research-design — historical methods, trend analysis, MROC, knowledge lifecycle, Universal Methods of Design router, research design document, report builder (load when the project needs formal design, design-method selection, or a final assembled report)
• source-evaluation — Wave-3 verification logic; mandatory pairing for every source
• academic-writing, report-and-proposal-craft, business-writing — output containers

Orchestration detail inputs

Orchestrator outputs

Evidence Produced

Capability Contract

Planning and review are read-only. Delegation is permitted only when the runtime and task authorise it; external contact, publishing, spending, or destructive actions require explicit authority.

Degraded Mode

Without delegation, network, or provider access, run waves sequentially and return the verified local result plus named gaps. Never collapse skipped verification into a pass.

Decision Rules

Worked Example

For a multi-region market question, assign one cohort per region, require the same comparison fields and evidence discipline, verify each checkpoint, then synthesise only reconciled claims.

Wave workflow

1. Define the question, cohorts, deadline, and evidence standard; stop if scope is unresolved.
2. Select research type, discipline, reading mode, and non-overlapping wave briefs.
3. Run each wave and verify its checkpoint before the next.
4. Dispatch gap-fill work where evidence conflicts or coverage is incomplete.
5. Recover from unavailable agents or providers by running sequentially and preserving unassessed gaps.

Wave anti-patterns

• Overlapping cohort briefs. Fix: assign exclusive boundaries.
• Synthesising before verification. Fix: gate each checkpoint.
• Omitting the evidence clause. Fix: include it verbatim.
• Treating no result as absence. Fix: report no source found.
• Hiding provider failure. Fix: mark it unassessed and recover sequentially.