Deep Researcher — the single-study research craft
Overview
deep-researcher runs one research study correctly, from question shape to
evidence report. The unit of work is a study folder in the research root: a plan, one
file per source with verbatim quotes, findings per hypothesis, and a final
answer-first report whose every number names its primary source. The method is
the measured one: recall is the binding constraint in retrieval, verification
budget lands at the merge, and every claim carries a rung, a lineage count and
a population label.
When to use
- A question feeds a decision, a published artifact, or a strategy, so a wrong answer costs more than the research.
- The answer needs more than one source type and a defensible trail: "research this question", "map the literature", "verify this claim properly", "write the evidence report", "investigate X before we commit".
When NOT to use: one-hop factual lookups (what, who, when, price), a single known URL to fetch, multi-day autonomous desk operations with worker fleets, or deliverables that must meet a formal systematic-review conduct standard (registration, dual-review attestations). This skill borrows that literature's measurement science; it never asserts compliance with it.
The precedence rule
The deepest loaded instruction governs its uniquely-claimed scope; every
scope: string is claimed by exactly one file in this tree; no leaf may widen
a prohibition declared by an ancestor — leaves narrow, never override upward.
The study pipeline
Run phases in order. Each leaf holds its full procedure; enter at the phase your study needs.
| Phase | You are doing | Leaf |
|---|---|---|
| plan | Sharpening the ask into a decision-backed question | subskills/scope-question.md |
| plan | Splitting the question into a round-by-round query plan | subskills/decompose-plan.md |
| retrieve | Grading sources on the trust ladder | subskills/source-triage.md |
| retrieve | Executing the keyless-first retrieval cascade | subskills/retrieval-cascade.md |
| verify | Verifying load-bearing claims and citations | subskills/verify-claims.md |
| verify | Attacking the draft with counter-evidence | subskills/adversarial-pass.md |
| synthesize | Merging evidence into findings | subskills/synthesize-findings.md |
| store | Writing claim cards for contested facts | subskills/evidence-cards.md |
| store | Keeping the claim-to-source-to-span ledger | subskills/evidence-ledger.md |
| stop | Deciding the study is done, with the reason stated | subskills/stop-rules.md |
| retrieve | Firing the grounded tier when its conditions hold | subskills/grounded-tier.md |
| store | Writing the study into the research root (the one folder that holds every study), incrementally | subskills/archive-writes.md |
| present | Delivering the answer-first evidence report | subskills/write-report.md |
Status & honest limits
Every leaf is status: seedling until a routing evaluation passes. The skill
encodes the procedure; it cannot force the harness to keep the discipline at
run time. Every operating rule the procedure needs is restated in plain words inside the
leaves; nothing here depends on an external rulebook.