phrazzld/research

Research

Retrieval-first research, multi-AI orchestration, and expert validation.

Execution Stance

You are the executive orchestrator.

• Keep query framing, source weighting, and final synthesis on the lead model.
• Delegate source retrieval and specialized analysis to focused subagents/tools.
• Run multi-source fanout in parallel for independent evidence streams.

Delegation Floor

Delegation floor applies: probe the roster first; dispatch two or more providers for substantive work; direct solo only for mechanical, emergency, user-forbidden, or fewer-than-two-providers cases. See harnesses/shared/AGENTS.md (Roster).

Local lane guidance: Use lanes with distinct sources, methods, or model families; web search, Thinktank, and provider CLIs are evidence inputs, not substitutes for lead synthesis. Native in-thread subagents do not count toward the roster floor.

Absorbed Skills

This skill consolidates: web-search, delegate, thinktank, introspect.

Routing

Explicit sub-capability (user names one)

If first argument matches a keyword, route directly to that reference:

| Keyword | Reference | |---------|-----------| | web-search, web-deep, web-news, web-docs | references/web-search.md | | delegate | references/delegate.md | | thinktank | references/thinktank.md | | introspect | references/introspect.md | | readwise | references/readwise.md | | xai | references/xai-search.md | | exemplars | references/exemplars.md |

No sub-capability (default): MANDATORY PARALLEL FANOUT

You MUST fan out to multiple sources. A single WebSearch is NOT research.

Research means gathering signal from many vantage points simultaneously. WebSearch alone is a lookup, not research. The whole point of this skill is multi-source triangulation.

REQUIRED: Launch ALL of these in a single message (parallel Agent/Bash calls). Missing any default source makes the run incomplete. Do not intentionally skip Thinktank or xAI/Grok in default mode:

1. Exa search — Bash: curl -s https://api.exa.ai/search -H "x-api-key: $EXA_API_KEY" ... See references/exa-tools.md for request format. WebSearch is fallback ONLY if curl fails.
2. Thinktank — Mandatory repo-aware bench for the default fanout. Run Thinktank from the repo you want it to inspect. Thinktank's workspace is context; --paths only orient the agents and do not replace the workspace root. If the query points at another repo, cd there before launch. Default fanout path: thinktank run research/quick --input "$QUERY" --output /tmp/thinktank-out --json --no-synthesis Deep path: thinktank research "$QUERY" --output /tmp/thinktank-out --json Add --paths ... when local files or directories should be pointed out as starting places. Before waiting, record the mode (quick or deep), output directory, and expected runtime band in your own notes. Default budgets: quick targets 60-180s with a hard cap of 300s; deep targets 3-8m with a hard cap of 900s. Thinktank launches a real multi-agent Pi bench, so even research/quick can take a few minutes. With --json, stdout stays reserved for the final envelope; a healthy run may look quiet until completion. Canonical inspection flow: thinktank runs show /tmp/thinktank-out and thinktank runs wait /tmp/thinktank-out. Inspect the output directory directly only when you need deeper detail: trace/events.jsonl, manifest.json, task.md, and prompts/ are the current partial-progress artifacts. Current limitation: if you stop a run early, completed agent reports may not exist yet. Treat that as a current Thinktank limitation, not as evidence that nothing happened. If Thinktank fails or times out, keep the Thinktank section and label it partial or failed with the output directory and observed error. That is an incomplete fanout, not a justified skip.
3. xAI / Grok — Bash: curl -s https://api.x.ai/v1/responses -H "Authorization: Bearer $XAI_API_KEY" ... Model MUST be grok-4.20-beta-latest-non-reasoning (only grok-4 supports tool use). Use xAI for grounded web retrieval, recency verification, contradiction checks, X-native discourse, and multimodal evidence. See references/xai-search.md for request format.
4. Codebase — Grep/Glob for what the project already does (skip only if query is unrelated to codebase)

Then produce a sourced report using the mandatory structure below.

Report Format (mandatory for all default fanout runs)

Every research report MUST have one labeled section per default source, followed by a synthesis. Include the section even when a provider returns partial output or fails; mark that status explicitly and carry the failure into the synthesis.

## Exa (neural search)
[Findings with inline URLs. What did Exa specifically surface?]

## xAI / Grok ([web_search | x_search | both])
[Findings with citations from response.citations. What did Grok surface?
For X Search: quotes or paraphrases from X posts, authors, dates.]

## Thinktank (Pi bench)
[What did the thinktank bench surface? Label this section `Thinktank (complete)`, `Thinktank (partial)`, or `Thinktank (failed)`. Note any disagreements between agents, the output directory, and the specific failure mode when applicable.]

## Codebase
[What relevant patterns, implementations, or prior art exist locally?
"None found" is a valid answer — write it explicitly.]

## Synthesis
[Consensus across sources. Conflicts or contradictions between them.
Recommendations grounded in the evidence above. Every claim cites a source.]

Discipline rule: if a section is missing, you failed the fanout. A failed provider still gets a section. A report that collapses all sources into one unlabeled blob has failed the fanout goal. Readers must be able to see what each tool contributed independently.

If Thinktank is still running when you need to respond, say that plainly. Name the output directory and summarize only the artifacts that actually exist so far. Do not present an incomplete run as a finished Thinktank result.

Narrow to a single source ONLY when:

• The user explicitly names one (e.g., "/research web-search [query]")
• It's a version/fact lookup (e.g., "what version is X?")

If you catch yourself about to return results from only WebSearch — STOP. That means you skipped the fanout. Go back and launch the other sources.

Use When

• Before implementing any system >200 LOC (reference architecture search)
• Before choosing a library, framework, or approach (current best practices)
• When training data may be stale (model releases, API changes, deprecations)
• When you need to verify a fact before asserting it
• When the user asks about something outside the codebase
• During /groom architecture critique (reference implementations)
• During /shape technical exploration (how others solve this)
• During /build understand step (existing patterns and examples)
• When another skill says "web search first" or "research before implementing"

Decision Framework

If you're about to assert something from training data that could be wrong, invoke /research web-search first. The cost of a search is negligible; the cost of hallucination is high.

This applies especially to:

• Model names and versions (stale within months)
• Library APIs and best practices (change with major versions)
• Pricing, availability, feature comparisons
• Security advisories, CVEs, deprecation notices

Provider Routing

| Query Type | Primary Provider | Why | |------------|-----------------|-----| | Repo-local architecture, tradeoffs, contradiction checks | Thinktank | Fresh multi-agent second voice against the workspace | | Code examples, reference implementations | Exa (code context) | Finds actual code, not blog posts | | Academic papers, formal specs | Exa (search) | Strong academic indexing | | Library/framework docs | Context7 | Semantic doc search | | Current events, model releases, fast recency checks | xAI Web Search | Grounded live web search with citations | | Social sentiment, public discourse, trending | xAI X Search | Native X/Twitter search and discourse retrieval | | Web pages with image/video analysis | xAI Web Search | Grounded web + multimodal | | General knowledge fallback | WebSearch / Brave | Broad coverage |

Default fanout uses Exa, Thinktank, xAI/Grok, and Codebase together. This routing table only tells you which source to lean on most heavily in synthesis; it does not authorize skipping the others. Exa remains strongest for code and technical indexing. xAI/Grok is broader than social pulse: use it for grounded web retrieval, recency verification, contradiction checks, and X-native discourse. Thinktank is the repo-aware second voice. See references/xai-search.md and references/thinktank.md for the concrete operator contract.

Anti-Patterns

• Asserting model versions from training data without searching
• Using WebSearch for code examples (Exa code context is better)
• Skipping research because "I'm pretty sure" (you're not)
• Research without citations (every claim needs a URL)