phrazzld/research

Research

Evidence-backed research. The lead owns framing, source weighting, synthesis, and residual uncertainty.

/research routes by capability, not vendor. Exa, xAI, Brave, Perplexity, Context7, Tavily, Firecrawl, browser agents, and provider lanes are acquisition tools behind a small set of evidence jobs: repo truth, docs lookup, web retrieval, code/context examples, extraction, social/discourse, recency verification, and synthesis.

Route

| Need | Load | |---|---| | broad research, comparison, architecture prior art, or discourse scan | references/default-fanout.md | | web-search, web-deep, web-news, web-docs | references/web-search.md | | Exa search/fetch/deep/MCP/code context | references/exa-tools.md | | extraction, site maps, crawls | references/extraction-tools.md | | delegate | references/delegate.md | | introspect | references/introspect.md | | readwise | references/readwise.md | | xai | references/xai-search.md | | exemplars | references/exemplars.md |

If the user names a sub-capability, load that reference. Otherwise use the default fanout for substantive research; narrow to one source only for explicit single-source requests or simple fact/version lookups.

Model selection/comparison is research, not memory: model facts rot in weeks. Start from the installed roster skill's references/model-provider-harness-index.md for current local facts, then verify against live sources — availability on the target platform, pricing, context, tool-calling support — and return a ranked recommendation with dates on every claim.

Contract

• Read the live repo first for repo facts.
• Use current external sources for drift-prone facts.
• Keep provider CLIs and web tools thin: launch, bound, record.
• Prefer acquisition surfaces in this order when available: MCP tool first, local CLI wrapper second, direct REST/API call third, built-in WebSearch last.
• Treat web search, extraction, X/social search, provider lanes, and local grep as evidence inputs, not substitutes for lead synthesis.
• Do not let synthesis stand in for retrieval. A grounded answer may summarize sources, but the source URLs/artifacts remain the proof.
• Separate source evidence from conclusions; cite URLs, files, commands, receipts, or artifacts for claims.
• Label skipped, failed, stale, in-flight, and partial sources.
• Report residual uncertainty instead of smoothing over missing evidence.

Delegation Judgment

delegate on judgment per the shared Roster contract: native subagents by default; add cross-model critics, roster providers, or sprite lanes (/sprites) only when they answer a distinct question. See harnesses/shared/AGENTS.md (Roster).

Local lane guidance: Use lanes with distinct sources, methods, or model families; web search and provider CLIs are evidence inputs, not substitutes for lead synthesis.

Completion Evidence

• Research objective and scope.
• Sources/tools queried and why.
• Provider lanes, receipt ids, accepted/rejected outputs, and failures.
• Claims tied to URLs, local files, commands, receipts, or artifacts.
• Source coverage gaps, stale facts, skipped tools, and residual risk.

Gotchas

• A single WebSearch is a lookup, not substantive research.
• Mandatory source structure belongs in references/default-fanout.md; keep vendor command recipes in tool references.