Deep-Research Prompts

Intermediate

Brief it like an analyst, not a search box

01 The Concept

Research modes (ChatGPT deep research, Claude research, Gemini Deep Research, Perplexity) run for minutes, browse dozens of sources, and return a report. The output quality is set almost entirely by the brief — a one-line question produces a generic explainer; a proper brief produces something you'd pay an analyst for.

A research brief has five parts: the decision the research serves ('choosing between X and Y for Z'), scope fences (time range, geography, what to exclude), source standards (what counts as evidence; what to distrust), the deliverable shape (sections, tables, comparison criteria), and honesty rules (mark thin evidence, separate fact from vendor claim, list what couldn't be verified).

02 Weak vs. Strong

EX 01Vendor evaluation research

Research customer-support platforms for a 40-person B2B SaaS (5-seat support team, ~2k tickets/mo, need Slack + Stripe integrations).

DECISION: shortlist 3 platforms to trial next month. SCOPE: mid-market tools only (not enterprise Zendesk tiers, not free-only tools); information from the last 18 months; pricing in USD. SOURCES: prioritize current pricing pages, public changelogs, and independent reviews with named authors; G2/Capterra aggregate scores are weak evidence; vendor comparison pages about competitors are claims to verify, not facts. DELIVERABLE:

Comparison table: seat price, per-ticket limits, Slack & Stripe integration depth (native/zapier/none), AI-agent features, migration path from Intercom
One paragraph per shortlisted tool: strongest fit signal + biggest risk
'Couldn't verify' list with what I should check in sales calls Mark any claim older than 12 months with its date — this market moves fast.

→ Why it works

Decision, fences, source hierarchy, exact table columns, and dated-claims discipline: the agent spends its browsing budget exactly where the trial decision needs it.

EX 02Market-entry scan

Research the public EV-charging market in Cambodia and Vietnam for a decision on whether a station-locator app should expand there in 2026.

ANSWER THESE, in order:

Installed public charger counts + 24-month growth, per country (cite primary/government sources where possible; date every figure)
Top 3 charging networks per country and whether they expose public APIs or app partnerships
Existing locator apps: names, approximate downloads, monetization
One regulatory factor that could change the picture within 18 months

RULES: separate 'measured' (installed base) from 'projected' (forecasts) — never blend them in one number; if Cambodia data is thin, say so explicitly rather than extrapolating from Vietnam; end with the 3 facts most likely to be wrong and how to check them locally.

→ Why it works

Numbered questions control report structure; measured-vs-projected separation kills the classic market-report sin; and thin-data honesty is demanded up front for exactly the country where data will be thin.

03 Key Points

01State the decision the research feeds — 'so I can decide X' reshapes everything the agent collects

02Fence the scope: time window, regions, market segment, and explicit exclusions

03Set source standards: 'prefer primary docs and independent tests; treat vendor blogs as claims, not evidence'

04Specify the deliverable: sections, a comparison table with named criteria, max length

05Demand epistemic honesty: confidence markers, 'couldn't verify' list, publication dates on key claims

06Iterate: skim the first report's sources, then rerun with tightened fences — two rounds beat one giant one

04 Model-Specific Notes

Claude's research mode follows structural instructions (numbered questions, table specs) very faithfully and is strong at 'couldn't verify' honesty when you ask for it explicitly.

05 For Your Role

Works for life decisions too: 'research these 3 neighborhoods for renting — decision criteria: commute under 30min to X, noise, flood risk. Separate data from real-estate-listing marketing.'

Next: RAG & Document Grounding →