deep-research
Use when conducting deep research on any topic that benefits from multiple behavioral perspectives, evidence gathering, and adversarial verification. Deploys math-optimized persona pairs with web sear
SKILL.mdplugins/agent-otto-founder/skills/deep-research/SKILL.md
GitHubConstellation Deep Research
Multi-perspective deep research engine that deploys math-optimized persona pairs to investigate any question. Each pair conducts independent research with web search, then debates findings internally. Captain synthesizes across all tracks.
Why This Beats Single-Perspective Research
| Perplexity / DeepSeek | Constellation Deep Research |
|---|---|
| 1 agent, 1 perspective | 3-4 optimized pairs, 6-8 behavioral lenses |
| Finds consensus | Finds productive disagreement |
| "Here's what the internet says" | "Here's what 4 teams found independently" |
| No blind spot detection | Blind spots ARE the product |
| Neutral tone | Each pair has behavioral signature |
| Research → report | Research → verdict with conviction scores |
| No adversarial check | Pairs stress-test each other's findings |
Commands
| Command | Pair Selection | Purpose |
|---|---|---|
/deep-research <question> |
Auto-select | General deep research |
/deep-research competitive <topic> |
Knowledge Bomb + Controlled Burn + Recon | Competitive intelligence |
/deep-research technical <topic> |
Mirror Match + Knowledge Bomb + Recon | Technical deep dive |
/deep-research market <topic> |
Knowledge Bomb + The Humanitarian + The Bridge | Market/people research |
Research Pair Roles
| Combo | Research Role |
|---|---|
| Knowledge Bomb (Analyzer + Persuader) | Core research. Finds evidence, packages for impact. ALWAYS included. |
| Mirror Match (Maverick + Specialist) | Innovation scan. Finds what nobody's tried, verifies feasibility. |
| Controlled Burn (Maverick + Guardian) | Risk assessment. Finds the bold angle, checks what could blow up. |
| Recon (Analyzer + Adapter) | Terrain mapping. Scouts adjacent spaces, stages context. |
| The Humanitarian (Altruist + Controller) | Impact analysis. Who benefits, who's harmed, what's the human cost. |
| The Bridge (Collaborator + Controller) | Stakeholder mapping. Who needs to know, how to translate findings. |
Execution Protocol
Step 1: Parse Query
Extract: research question, depth (quick/standard/deep), scope, pair override.
Step 2: Select Research Pairs
Default: Knowledge Bomb (always) + 2-3 pairs based on query type.
Step 3: Write Research Brief
~200 token brief: question, known context, what we're looking for, scope boundaries.
Step 4: Dispatch Research Pairs (Parallel)
Each pair subagent gets:
- Both persona profiles
- The research brief
- Access to web search and web fetch tools
- Instructions to conduct independent research AND debate findings
IMPORTANT: All pairs dispatch in a SINGLE message (parallel).
Each subagent prompt:
You are a RESEARCH PAIR: {COMBO_NAME} ({PERSONA_A} + {PERSONA_B}).
Pair Math: Balance {BALANCE}, Energy {ENERGY}.
## Research Question
{BRIEF + QUESTION}
## Your Research Task
1. Use WebSearch and WebFetch to investigate the question
2. Gather 3-5 high-quality sources
3. Debate findings as {PERSONA_A} and {PERSONA_B}
4. Deliver a joint research verdict
## Format
### Sources Found
1. [Source title](URL) — key finding
2. [Source title](URL) — key finding
3. ...
### Research Discussion
{PERSONA_A}: [initial findings, in character]
{PERSONA_B}: [challenges, adds rigor, in character]
{PERSONA_A}: [responds]
{PERSONA_B}: [final verification]
### {COMBO_NAME} RESEARCH VERDICT
- Key finding: [one clear statement backed by evidence]
- Evidence strength: [strong/moderate/weak]
- Gaps in evidence: [what we couldn't find or verify]
- Recommended follow-up: [specific next research step]
- Conviction: [0-10]
Step 5: Collect Research Tracks
Parse each pair's findings, sources, evidence strength, and verdict.
Step 6: Cross-Reference Sources
- Flag contradictions between pairs
- Identify sources cited by multiple pairs (high confidence)
- Note gaps where no pair found evidence
Step 7: Captain Synthesis
Synthesize across all research tracks:
- Consensus findings — what all pairs agree on, with sources
- Contested findings — where pairs disagree, with evidence for each side
- Blind spots — what no pair investigated
- Source quality map — which sources are most/least reliable
- Actionable conclusions — ranked by evidence strength and conviction
Step 8: Persist Report
Write to ~/Desktop/Airlock/repos/airlock-coordination/reports/
Include: format: deep_research, pairs:, sources:, evidence_map:.
Step 9: Render Report
══════════════════════════════════════════════════
CONSTELLATION DEEP RESEARCH
Query: "{question}"
Date: {date} | Research Pairs: {count}
Sources: {total_sources} | Avg Conviction: {N}
══════════════════════════════════════════════════
RESEARCH BRIEF
{brief}
──────────────────────────────────────────────────
{COMBO_NAME} — Research Track
Sources: {N} | Evidence: {strong/moderate/weak} | Conviction: {N}
Key Finding: {finding}
Sources:
1. {source + key takeaway}
2. {source + key takeaway}
Discussion Highlights:
{PERSONA_A}: {key point}
{PERSONA_B}: {key challenge/addition}
Gaps: {what couldn't be verified}
──────────────────────────────────────────────────
(repeat for each pair)
══════════════════════════════════════════════════
EVIDENCE MAP
Strong (cited by 2+ pairs): {findings}
Moderate (single pair, good source): {findings}
Weak (single pair, limited source): {findings}
Contested: {finding} — {pair_A} says X, {pair_B} says Y
══════════════════════════════════════════════════
SYNTHESIS
{Captain's synthesis with ranked conclusions}
Recommended Actions:
1. {action + evidence strength}
2. {action + evidence strength}
══════════════════════════════════════════════════
Cost Comparison
| Service | Cost | Perspectives | Adversarial Check |
|---|---|---|---|
| Perplexity Pro | $20/mo flat | 1 | No |
| Deep Research (3 pairs) | ~$0.20/query | 6 | Yes, per-pair |
| Deep Research (4 pairs) | ~$0.30/query | 8 | Yes, per-pair |
100 deep research queries/month = ~$25. Comparable to Perplexity Pro but with behavioral optimization, adversarial tension, and conviction scoring.
The Moat
Nobody else has the sovereign balance math to know WHICH pairs to deploy for WHICH type of question. The 136-pair matrix ranked by behavioral mathematics is not a prompt trick — it's a research methodology grounded in PI behavioral science. Perplexity can't replicate behavioral tension because they don't have the framework underneath.