Deep Research & Synthesis

Produce research a skeptical expert would trust: every load-bearing claim sourced, reliability labeled, conflicts surfaced, and the synthesis honest about what's known vs. estimated.

Principles

1. Scope before searching. Pin down the question, audience, deliverable, and depth. For research, start an initial recon search immediately rather than gating everything on a clarifying question.
2. Ground in primary/first-party data first. If the subject owns data (analytics, a repo, internal docs), pull it before web opinion.
3. Fan out, don't trickle. Decompose into independent streams; research in parallel (parallel subagents when available).
4. Demand provenance. Every statistic gets a source name, URL, and year.
5. Tier reliability. Primary/academic/first-party vs directional/vendor; use directional figures for direction only (references/source-tiering.md).
6. Verify the load-bearing claims against primary sources; drop or label what doesn't survive.
7. Reflect and stop deliberately. After each pass: what did I find, what's missing, is coverage enough? Stop when saturated (references/research-rigor.md).
8. Surface conflicts; never fabricate.

Workflow

1. Scope & clarify, then write a one-paragraph research brief (constraints stated, unknowns marked open).
2. Recon the subject and any first-party data to sharpen the stream plan.
3. Plan 3–5 independent streams mapped to eventual report sections.
4. Fan out with precise, self-contained briefs (references/subagent-briefs.md); pull interactive/first-party data yourself.
5. Collect with discipline (URLs + years; concrete examples; note source type).
6. Tier, score, and verify (credibility scoring + primary-over-secondary in references/research-rigor.md); re-fetch primaries for load-bearing stats.
7. Coverage check, then synthesize — lead with the conclusion, weave inline citations, show conflicts, end with a tiered source list. Hand to report-information-design for a polished deliverable.

Reference map

• references/research-playbook.md — recon→plan→fan-out→verify→synthesize loop in detail.
• references/research-rigor.md — query expansion, the reflection/gap loop, stop conditions, credibility scoring, self-contained briefs + verbatim-return contract (from gpt-researcher / STORM / open_deep_research).
• references/source-tiering.md — the reliability rubric, confidence phrasing, handling conflicts.
• references/subagent-briefs.md — the parallel-research brief template + worked examples.

Engagement context

At the start of a task, check for the engagement brief (default ./cro-engagement.md, created by the getting-started skill / /kickoff). If it exists, read it first and tailor scope, audience, and depth to it. If it's missing and the work is more than a one-off, offer to capture one via getting-started.

Related skills

report-information-design (turn the synthesis into a report), ecommerce-cro-audit (a frequent consumer of this skill's evidence). Agents: deep-researcher runs a single stream; cro-verifier adversarially checks the result.

Honesty guardrails

Only cite URLs actually opened in this session, and run python3 ${CLAUDE_PLUGIN_ROOT}/skills/report-information-design/scripts/check_citations.py <synthesis-file> on the synthesis before handing it on (LLM research tools fabricate or break a meaningful share of links). For a load-bearing claim surfaced only by AI search, find a second independent source. Label vendor/self-interested figures directional; distinguish "verified against primary" from "widely cited but untraced"; prefer recent sources for fast-moving topics; include a short source-reliability appendix so the reader can defend the work.