deep-research — fan-out and adversarial verifyCapability

When a question needs many sources cross-checked — LASIK vs ICL, Innovation Director comp bands, therapist options — stop hand-searching one tab at a time. The /deep-research harness fans out parallel searches, reads the sources, tries to disprove its own findings, and hands you a cited report.

Hand-searching is sequential and trusting: you open a tab, skim, open another, and quietly believe whatever the most confident page said. It's fine for 'what time does the store close' and dangerous for 'which eye surgery is right for my eyes' — the failure mode is a plausible single source that's wrong, outdated, or selling something.

A research harness flips both problems. Fan-out: it runs many searches at once from different angles, so no single page dominates. Adversarial verify: before a claim makes the final report, a separate pass actively tries to refute it — the same instinct as your own CLAUDE.md automation rule 'verify the side-effect, not the trigger', applied to facts instead of cron jobs. What survives the refutation attempt is what you read.

In your workspace this is the /deep-research skill. It's the heavyweight cousin of just asking me a question: slower and more thorough, it ends in a synthesized report with citations you can click. Reach for it when being wrong is expensive or when you'd otherwise open ten tabs — not for quick lookups, where it's overkill.

Worked example

Turn the LASIK-vs-ICL task into a real research run. The skill wants a specific question, not a topic — so narrow it first with your own facts (age, prescription, budget, DC location):

/deep-research Compare LASIK vs ICL for a 30-something with moderate myopia in
Washington DC: cost ranges at reputable DC-area clinics, recovery timelines,
candidacy criteria (corneal thickness, dry eye), and the real risk profiles of
each. Cite clinical sources, not clinic marketing pages.

• A specific question beats a topic. 'LASIK vs ICL' returns mush; the version above tells the harness your age band, prescription severity, region, and the four exact dimensions your Things task already lists (cost, recovery, candidacy, risk).
• The last sentence is a source-quality instruction. Saying 'cite clinical sources, not clinic marketing' steers the adversarial pass to distrust pages whose job is to sell you surgery — the single biggest trap in medical search.
• The skill's own description says: if the question is underspecified it will ask 2-3 clarifying questions first. Answer them — that's the harness narrowing scope, and it's what makes the report yours instead of generic.

Gotchas

• Underspecified questions waste the run. 'What car should I buy' with no budget/use-case gets you a generic essay — the skill even calls this out and will ask first. Front-load your constraints.
• It's not for quick facts. A single lookup ('Resend pricing tier') is faster asked directly — the harness's fan-out is wasted effort there.
• Citations are the deliverable, not decoration. A confident report with no clickable sources is exactly the failure deep-research exists to prevent — if you can't verify a claim, treat it as unverified.