research

name	research
description	Execute a research task using compositional workflow planning. Characterizes the task, selects an approach from a library of research strategies, executes it, and self-evaluates output quality.
argument-hint	["research question or topic"]
user-invocable	true
allowed-tools	Read, Glob, Grep, Bash, WebFetch, WebSearch

Research this topic: $ARGUMENTS

Workflow

Phase 1: Task Characterization

Rate the research task on each dimension (1-5):

Dimension	1 (Low)	5 (High)
Scope	Point question (single fact)	Frontier mapping (state of entire field)
Domain structure	Single field, established methods	Distant cross-domain analogy
Evidence type	Empirical data, measurements	Theoretical arguments, first-principles
Time horizon	What's true right now	What could become true (speculative)
Fidelity	Ballpark / directional	Rigorous / publication-grade

Phase 1b: Workflow Library Query

Query the MAP-Elites workflow library for strategies matching this task's coordinates:

sqlite3 research-workflows/workflows.db "SELECT id, name, archetype, fitness_score FROM workflows WHERE status IN ('active', 'seed') ORDER BY fitness_score DESC LIMIT 10;"

For more targeted retrieval, filter by behavioral region:

sqlite3 research-workflows/workflows.db "SELECT w.id, w.name, ws.name as step_name, ws.description, ws.rationale FROM workflows w JOIN workflow_steps ws ON w.id = ws.workflow_id WHERE w.archetype = '<archetype>' AND w.status IN ('active', 'seed') ORDER BY w.fitness_score DESC, ws.step_order;"

Use retrieved workflows to inform Phase 2 strategy selection. Prefer workflows with high fitness scores and usage counts.

Phase 2: Strategy Selection

Based on the characterization and library query, select the most appropriate research archetype:

Exploratory (high scope, moderate fidelity)

• Quick landscape scan — 15-minute overview of a new area
• Deep literature review — comprehensive survey of a mature field
• Trend detection — what's gaining momentum
• White space identification — what isn't being worked on that should be

Confirmatory (low scope, high fidelity)

• Fact-checking pipeline — verify specific claims with primary sources
• Consensus mapping — what do experts agree/disagree on
• Replication check — has this finding held up under scrutiny

Analytical (moderate scope, high evidence)

• Compare and contrast — systematic comparison of N approaches
• Root cause analysis — why did X happen / why doesn't X work
• Cost-benefit analysis — should we do X given tradeoffs
• Forecasting — given current trends, what's likely

Generative (high domain structure, moderate time horizon)

• Cross-domain transfer — find solutions from field B for problems in field A
• First-principles derivation — reason from fundamentals, not literature
• Synthesis — combine N existing ideas into a novel framework
• Adversarial stress-test — find the strongest objections to X

Applied (low time horizon, high fidelity)

• Technical feasibility assessment — can X be built with current tools
• Build-vs-buy analysis — build or use existing solution
• Implementation planning — how to execute X given constraints

Phase 3: Compositional Planning

Don't just pick one strategy. Identify which sub-techniques from multiple strategies suit this specific task:

1. Select 2-3 relevant archetypes from Phase 2
2. For each, identify the steps that apply to this task
3. Note why each step exists (what it optimizes for)
4. Compose a novel workflow combining the best elements
5. Record which elements were borrowed from which archetype

Phase 4: Execution

Execute the composed workflow using available tools:

• Exa Search API (via manage_websets) — exa.search (instant web search with category/date/domain filters), exa.findSimilar (find pages similar to a URL), exa.getContents (extract text/highlights/summary from URLs), exa.answer (question answering with citations)
• Exa Task Workflows (via manage_websets tasks.create) — retrieval.searchAndRead (search + read pages), retrieval.expandAndCollect (search + findSimilar expansion + dedup), retrieval.verifiedAnswer (answer + independent validation)
• Exa Webset Workflows (via manage_websets tasks.create) — lifecycle.harvest, convergent.search, adversarial.verify, qd.winnow, research.deep, research.verifiedCollection
• WebSearch — broad web queries
• WebFetch — deep read of specific URLs
• Firecrawl (firecrawl_scrape, firecrawl_search) — structured web scraping
• GitHub — repository search, code search, trending
• Read/Glob/Grep — local codebase and documentation

Phase 5: Self-Evaluation

Score your output on each dimension (0-1):

1. Coherence — Does it tell a consistent story? Internal contradictions?
2. Grounding — Are claims supported by retrieved evidence?
3. Compression — Can key findings be stated concisely? (If not, thesis may be unclear)
4. Surprise — Did anything non-obvious surface? (Confirming the known has low value)
5. Actionability — Does the output enable a decision or next step?

Report the scores honestly. Low scores on surprise or actionability are useful signals — they indicate the research question may need reframing.

Phase 5b: Library Update

Log this execution and update the workflow library:

# Log the execution
sqlite3 research-workflows/workflows.db "INSERT INTO executions (task_description, task_scope, task_domain_structure, task_evidence_type, task_time_horizon, task_fidelity, score_coherence, score_grounding, score_compression, score_surprise, score_actionability, score_composite) VALUES ('<description>', <scope>, <domain>, <evidence>, <horizon>, <fidelity>, <coherence>, <grounding>, <compression>, <surprise>, <actionability>, <composite>);"

If composite score exceeds the current occupant of this behavioral region, the composed workflow becomes a candidate for library insertion.

Phase 6: Output

## Research: [Topic]

### Characterization
Scope: [1-5] | Domain: [1-5] | Evidence: [1-5] | Horizon: [1-5] | Fidelity: [1-5]

### Strategy
[Which archetypes were combined and why]

### Findings
[Structured findings — use headers, tables, or lists as appropriate]

### Quality Assessment
Coherence: [0-1] | Grounding: [0-1] | Compression: [0-1] | Surprise: [0-1] | Actionability: [0-1]

### Key Takeaway
[Single paragraph — the compressed thesis]

### Recommended Next Steps
[What to do with these findings]