researcher-deep-dive

researcher-deep-dive

Work class: research (authoritative, non-downgradable). Any lane pulling this skill is purpose: research and resolves effective_complexity: hard / tier: powerful (research_always_hard — the frozen guild.model_policy.v2 floor). The floor is transitive: every handoff, advisor consult, retry, resume, or sub-dispatch descending from this lane inherits it and cannot downgrade it. Handoff receipts record purpose: research, effective_complexity: hard, tier: powerful, forced_floor_reason: research_always_hard.

Evidence is the source list — every non-trivial claim cites a source with a URL and a confidence label.

What you do

Synthesize what is actually known about a topic from multiple sources, separate claims by confidence, and surface the open questions the team still has to answer. The brief is scannable in five minutes and auditable in thirty.

• Pull from at least 3–5 distinct sources; note primary vs secondary and date of each.
• Group findings into claims with a confidence tag: well-established, contested, emerging, speculative.
• Quote exactly when a claim is load-bearing — paraphrase only when safe.
• Distinguish what is known from what is inferred; flag inferences.
• End with explicit open questions the sources do not answer.
• Include a TL;DR at the top (3–5 bullets) the reader can cite directly.

Output shape

Markdown file at .guild/runs/<run-id>/research/<slug>.md:

1. TL;DR — 3–5 bullets.
2. Scope & framing — what question this brief answers, what it deliberately doesn't.
3. Key findings — each finding tagged with confidence and linked sources.
4. Contested / open questions — what the sources disagree about.
5. Sources — numbered list: title · author · date · URL · primary/secondary.
6. Methodology note — one paragraph on how the brief was assembled (search strategy, exclusions).

Anti-patterns

• Unsourced claims — "it's widely known that…" is not a source.
• One-source synthesis — echoing a single article dressed up as research.
• Editorial opinion as finding — if it's your take, label it and put it in a clearly-marked section.
• No confidence calibration — stapling "research shows" onto contested claims flattens signal.
• No summary — a 20-page brief without a TL;DR won't be read.
• Cherry-picked quotes — quoting the 10% of a source that agrees with a preferred answer.

Handoff

Return the brief path to the invoking researcher specialist. If the brief names specific options to compare, the researcher chains into researcher-comparison-table; if it names a specific paper to dig into, into researcher-paper-digest. This skill does not dispatch.