Deep Research

<identity> Deep Research Skill — Multi-step autonomous research methodology for structured, evidence-based investigation and synthesis. </identity> <capabilities> - 5-phase autonomous research workflow (Scope → Search → Synthesis → Validation → Report) - Multi-source evidence gathering with cross-validation - Structured output with citations and confidence ratings - Gap detection and iterative depth expansion - Actionable recommendation generation </capabilities>

When to Use

Task requires investigation across multiple sources before drawing a conclusion
Evidence-based answers needed, not pattern-matched guesses
Question spans multiple domains and needs synthesis
Accuracy of findings is critical to downstream decisions

Trigger phrases: "research", "investigate", "deep dive", "find evidence for", "comprehensive analysis of"

Iron Law

Never synthesize without reading sources. Every claim in the final report must trace to a source retrieved during Phase 2.

5-Phase Process

Phase 1: Scope — Restate as one research question. Define 3–5 sub-questions, success criteria, constraints, and non-goals. Cannot proceed without a written scope block.

Phase 2: Multi-Source Search — Run 3+ WebSearch queries. WebFetch each source URL. pnpm search:code for prior-art. Maintain evidence log: URL, claim, confidence (HIGH/MEDIUM/LOW), sub-question. Each sub-question needs ≥2 independent sources.

Phase 3: Synthesis — Group evidence by sub-question. CONSENSUS (2+ sources) → HIGH; CONFLICT → LOW, flag for user. Build synthesis matrix. Label unsupported claims [UNVERIFIED].

Phase 4: Validation

Source freshness — appropriate time period?
Contradiction check — resolve or flag CONFLICT items.
Assumption audit — 3 assumptions stated explicitly?
Scope alignment — each finding maps to a Phase 1 sub-question?
Confidence calibration — downgrade single-source HIGH ratings.

Phase 5: Report — Save to .claude/context/reports/backend/<topic>-research-<YYYY-MM-DD>.md with sections: Executive Summary, Findings (finding + confidence + evidence per sub-question), Conflicts, Recommended Actions, Knowledge Gaps, Sources.

Tool Usage

Tool	Phase	Purpose
`WebSearch`	2A	Broad source discovery
`WebFetch`	2B	Extract content from URLs
`Read` / `Grep`	2C	Search codebase for prior art
`Write`	5	Save report

Output Locations

Reports: .claude/context/reports/backend/<topic>-research-<YYYY-MM-DD>.md
Evidence logs (temp): .claude/context/tmp/research-evidence-<YYYY-MM-DD>.md

Search Protocol

pnpm search:code "<query>" — primary intent-based search
Skill({ skill: 'ripgrep' }) — exact keyword/regex matches
WebSearch / WebFetch — external sources
Grep — fallback for single-file targeted checks

Iron Laws

NEVER proceed to Phase 3 without ≥2 sources per sub-question.
ALWAYS save evidence log to tmp/ before Phase 3 — context resets lose unsaved evidence.
ALWAYS label confidence explicitly — never present LOW-confidence as conclusion.
NEVER paper over gaps — unknown answers must be flagged, not omitted.
ALWAYS verify report file exists before marking task complete.

Anti-Patterns

No scope block before searching → unfocused evidence, unsynthesizable
Single source per claim → cross-validate with ≥2 independent sources
Synthesizing while searching → introduces confirmation bias; complete Phase 2 first
Omitting CONFLICT evidence → flag all conflicts explicitly in report

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md

Check for: prior reports on same topic (avoid duplicates), unreliable sources in issues.md, relevant ADRs in decisions.md.

After completing:

New research pattern -> .claude/context/memory/learnings.md
Unreliable source identified -> .claude/context/memory/issues.md
Methodology decision -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: Save evidence log and partial findings before Phase 3.

deep-research