Deep Research — Asymmetric Research Squad

You are the orchestrator: deploy eight distinct research personas, run a crucible analysis, generate emergent insights, then synthesize actionable intelligence. Preserve contradictions and uncertainty; do not replace missing research with guesses.

Phase 0: Define and Set Up

Parse $ARGUMENTS: flags first, remaining text is the required research subject.

Default: standalone parallel sub-agents.

• --team: Claude Code only; opt into a shared team of 14 agents across three batches. Cursor and Codex lack team tools.
• --output-dir "...": output directory; default research/[sanitized-subject].
• --dry-run: display the plan without running checks, creating files, or dispatching.

Missing subject: abort with usage and an example:

/ycc:deep-research [--team] [--output-dir "..."] [--dry-run] <research-subject>
/ycc:deep-research --dry-run "AI model deployment strategies"

Require a focused topic. Sanitize the default directory name: lowercase, replace spaces with hyphens, remove characters other than alphanumerics and hyphens. For --team, require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; otherwise abort with --team requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Retry without --team or intentionally enable agent teams.

Dry run

Display the subject, output path, four phases (8 personas, 2 crucible agents, 4 strategic agents, report synthesis), and this artifact tree: objective.md, persona-findings/{historian,contrarian,analogist,systems-thinker,journalist,archaeologist,futurist,negative-space}.md, synthesis/{crucible-analysis,contradiction-mapping,tension-mapping,pattern-recognition,negative-space,innovation}.md, evidence/verification-log.md, report.md. STOP: do not run prerequisite scripts, dispatch agents, create directories, or write artifacts.

Agent-team dry-run additions

With --team, also show drpr-<sanitized-subject>, 14 teammates, and dependencies: Batch 2 blocked by all Batch 1 tasks; Batch 3 blocked by both Batch 2 tasks. Do not call TaskCreate, Agent, or SendMessage in dry-run mode.

Infrastructure and objective (not dry run)

Run scripts/check-prerequisites.sh when preparing the output directory, before creating anything:

${CLAUDE_PLUGIN_ROOT}/skills/deep-research/scripts/check-prerequisites.sh "$OUTPUT_DIR"
mkdir -p "$OUTPUT_DIR/persona-findings" "$OUTPUT_DIR/synthesis" "$OUTPUT_DIR/evidence"

Write $OUTPUT_DIR/objective.md with 3–5 research questions, success criteria, evidence standards, perspectives to cover, and 2–3 subject-specific bias risks. Use the phase checklist below as the success criteria. Give every agent the absolute objective and output paths, subject, and any user scope constraints.

References and Routing

Resolve paths against ${CLAUDE_PLUGIN_ROOT}/skills/deep-research/. Read only the reference needed for the current phase; report any unreadable file. Do not read or paste all persona prompts into orchestrator context. Each persona agent reads its own file. The table gives exact paths and read conditions.

Reference	Read when
references/persona-historical.md	Historian agent starts Phase 1: evolution, failures, forgotten alternatives.
references/persona-contrarian.md	Contrarian agent starts Phase 1: disconfirming evidence, critiques, failures.
references/persona-analogical.md	Analogist agent starts Phase 1: structural parallels across domains.
references/persona-systems.md	Systems Thinker agent starts Phase 1: feedback, stakeholders, causal chains.
references/persona-journalistic.md	Journalist agent starts Phase 1: current state, players, dated developments.
references/persona-archaeological.md	Archaeologist agent starts Phase 1: obsolete solutions from 10–50 years ago.
references/persona-futurist.md	Futurist agent starts Phase 1: patents, speculative research, 2030+ forecasts.
references/persona-negative-space.md	Negative Space Explorer agent starts Phase 1: omissions, barriers, gaps.
references/analysis-prompts.md	Crucible or strategic agents start Phases 2–3: use their named mandate.
references/verification-log-template.md	Orchestrator verifies critical claims after Phase 2: preserve claimants, sources, verification status, and competing claims with resolution.
references/completion-summary-template.md	Orchestrator reports completion after Phase 4: formatted summary with separate counts for all four strategic analyses.
references/synthesis-template.md	Orchestrator starts Phase 4: report structure and writing checks.

Dispatch Contract (All Batches)

Standalone (default)

See ${CLAUDE_PLUGIN_ROOT}/skills/_shared/references/standalone-dispatch.md for the standalone-dispatch contract (Task vs Agent+team, anti-patterns, backstop policy). Spawn each batch in one message. Each agent writes its declared artifact. When a result comes back empty or cut short, re-read the matching artifact from disk and treat it as authoritative before declaring that dispatch failed. Record a single failure as a gap in its artifact; adapt later phases. If at least half a batch fails, abort and report the gaps instead of fabricating results.

Agent-team dispatch (--team only)

Read ${CLAUDE_PLUGIN_ROOT}/skills/_shared/references/agent-team-dispatch.md when --team is used. Do not mix standalone Agent calls with team Agent calls.

1. Sanitize subject for team name: lowercase; replace [^a-z0-9-] with -; collapse hyphens, trim edges, truncate to 20 characters; fallback untitled.
2. Confirm the implicit session team (flag CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1); abort if absent.
3. TaskCreate all 14 roster tasks up front, using teammate name as subject prefix. Capture real task IDs. TaskUpdate both Batch 2 tasks with all eight Batch 1 IDs in addBlockedBy; each Batch 3 task gets both Batch 2 IDs. On registration failure, shut down any spawned teammates and abort.
4. Spawn each batch in one message with multiple Agent calls. Every call carries unique name from the roster, subagent_type, and its mandate.
5. Confirm that batch's tasks are completed via TaskList before consuming outputs. Send shutdown_request via SendMessage to every teammate in that batch before advancing. Keep the team through Phases 1–3.
6. After final batch shutdown, confirm all teammates have shut down before Phase 4 synthesis.

Failure policy: record a single failure as a gap in its artifact and adapt later phases. If at least half a batch fails, abort; in team mode shut down active teammates. Suggest retry without --team. On user abort, shut down all active teammates. Never leave a teammate live after exit.

Phase 1: Asymmetric Persona Deployment (Batch 1)

All eight use research-specialist. Output paths are relative to $OUTPUT_DIR.

Persona	Teammate name	Persona file	Output file
Historian	historian	references/persona-historical.md	persona-findings/historian.md
Contrarian	contrarian	references/persona-contrarian.md	persona-findings/contrarian.md
Analogist	analogist	references/persona-analogical.md	persona-findings/analogist.md
Systems Thinker	systems-thinker	references/persona-systems.md	persona-findings/systems-thinker.md
Journalist	journalist	references/persona-journalistic.md	persona-findings/journalist.md
Archaeologist	archaeologist	references/persona-archaeological.md	persona-findings/archaeologist.md
Futurist	futurist	references/persona-futurist.md	persona-findings/futurist.md
Negative Space Explorer	negative-space-explorer	references/persona-negative-space.md	persona-findings/negative-space.md

CRITICAL: Deploy all 8 persona agents in a SINGLE message with MULTIPLE Agent tool calls (standalone mode only).

In --team mode, use multiple Agent calls in one message instead, following the team dispatch contract.

Even for a small topic, pass each resolved absolute persona file path, not its text. Each prompt follows this shape, with concrete values:

Read <absolute plugin root>/skills/deep-research/references/persona-<name>.md first.
RESEARCH_SUBJECT=<subject>
OUTPUT_DIR=<absolute output directory>
Read <output directory>/objective.md. Substitute these variables in your brief.
Use your distinct search strategy; cite evidence and confidence, record uncertainty.
Write <absolute output file> and return findings. User constraints: <constraints>.

Require 8–10 diverse queries per persona; brief findings do not mean skipping personas. Track completion of each persona and its artifact. Wait for all eight (or explicit failure records) before Phase 2.

In team mode, shut down Batch 1.

Phase 2: The Crucible (Batch 2)

Read all eight persona artifacts and objective.

CRITICAL: Deploy both analysis agents in a SINGLE message with MULTIPLE Agent tool calls (standalone mode only).

In --team mode, use multiple Agent calls in one message instead, following the team dispatch contract.

Pass the references/analysis-prompts.md resolved path and the named mandate, subject, output directory, and input artifact paths. Agents read their own section.

Agent / teammate name	Subagent type	Output
ACH Analyst / ach-analyst	general-purpose	synthesis/crucible-analysis.md
Contradiction Mapper / contradiction-mapper	codebase-research-analyst	synthesis/contradiction-mapping.md

ACH: generate 5+ mutually exclusive hypotheses; weigh disconfirming evidence. Contradiction Mapper: preserve disagreements, assess sources, identify patterns. After completion, triangulate critical claims: fetch primary sources, document confirmations, partial confirmations, confidence, competing evidence, and unresolved contradictions in $OUTPUT_DIR/evidence/verification-log.md. Read references/verification-log-template.md when writing this log and preserve its fields. Claims repeated by multiple personas still need verification, not a vote.

In team mode, shut down Batch 2.

Phase 3: Emergent Insights (Batch 3)

CRITICAL: Deploy all 4 strategic analysis agents in a SINGLE message with MULTIPLE Agent tool calls (standalone mode only).

In --team mode, use multiple Agent calls in one message instead, following the team dispatch contract.

Give each the resolved references/analysis-prompts.md path, named mandate, objective, all persona findings, both crucible outputs, and verification log. Require the following artifacts:

Agent / teammate name	Subagent type	Output
Tension Mapper / tension-mapper	codebase-research-analyst	synthesis/tension-mapping.md
Pattern Recognizer / pattern-recognizer	general-purpose	synthesis/pattern-recognition.md
Negative Space Analyst / negative-space-analyst	codebase-research-analyst	synthesis/negative-space.md
Innovation Agent / innovation-agent	general-purpose	synthesis/innovation.md

Keep negative-space-analyst distinct from Phase 1's negative-space-explorer. Consume all four results (or gap records).

In team mode, shut down Batch 3 teammates now, not after reporting.

Phase 4: Strategic Synthesis (Orchestrator Only)

Read all artifacts: eight personas, two crucible analyses, four strategic analyses, and verification log. Read references/synthesis-template.md when composing the final report. Write $OUTPUT_DIR/report.md using its structure.

Synthesis rules

• Organize by insight/theme, not persona; combine related findings without losing distinct perspectives. Executive synthesis is information-dense, not a recap.
• Highlight contradictions; capture both claims, evidence, extra investigation, and which is better supported. Retain unresolved tensions and confidence ratings.
• Apply evidence hierarchy: primary (original research, official records, first-hand accounts) > secondary (expert/peer-reviewed analysis) > synthetic (aggregated media)

  speculative (predictions, opinions, unverified claims).

• Separate facts, analogies, and forecasts. Preserve uncertainty and missing evidence.
• Focus on actionable implications: second/third-order effects, stakeholders, leverage points, and explicit further-research questions.

Run scripts/validate-research.sh when report and artifacts are ready:

${CLAUDE_PLUGIN_ROOT}/skills/deep-research/scripts/validate-research.sh "$OUTPUT_DIR"

Fix missing/empty artifacts, weak citation coverage, and format issues. Report actual validation result and any remaining gaps; do not call a failed run complete. Read references/completion-summary-template.md when presenting the formatted completion summary. Preserve its per-persona counts and separate Phase 3 counts: tensions, patterns, research gaps, and novel hypotheses. Report actual counts, discoveries, unresolved disagreements, evidence quality, and artifact paths; do not collapse the four strategic analyses into a generic emergent-insights count.

Method and Phase Checklist

Query variation: SCAMPER

Substitute X with Y; Combine X and Y; Adapt another domain; Modify scale/scope; Put to other uses; Eliminate X; Reverse assumptions. Use 8–10 well-crafted, diverse queries rather than generic volume. Keep each persona's mandate distinct.

Cover past (10–50 years), recent past (1–10 years), present, near future (1–5 years), and far future (5+ years). Sources must be authoritative and dated/current where relevant.

Phase 1

• All 8 personas deployed with distinct search strategies
• 8–10 queries per persona executed.
• Findings cite primary sources, confidence, uncertainties, and contradictions.
• Historical, cross-domain, systemic, present, future, and missing perspectives covered.

Phase 2

• ACH has 5+ mutually exclusive hypotheses and actively seeks disconfirmation.
• Cross-persona contradictions and patterns mapped, not smoothed over.
• Primary-source verification log maintained; major claims have citations.

Phase 3

• Tension mapping identifies maximum disagreement.
• Pattern recognition finds unexpected connections.
• Negative space documents unanswered questions.
• Innovation combines perspectives into novel, evidence-rated hypotheses.

Phase 4

• Executive synthesis is dense intelligence, not a persona-by-persona summary.
• Evidence portfolio includes confidence and preserves distinct viewpoints.
• Strategic implications cover second/third-order effects and temporal range.
• Research gaps and validation result explicitly reported.
• Teammate shutdown completed via SendMessage if applicable.

Corrective Actions

Similar findings: revisit individual mandates and SCAMPER strategies. Speculation outweighing evidence: fetch primary sources. Shallow research: inspect query logs. Missing contradictions: seek disconfirming evidence rather than smoothing tensions. Summary-only report: reread artifacts and synthesize insights. Missing gaps: revisit both negative-space roles. Never silently invent evidence to fill a failed artifact.