Prompting deep research tools to deliver tables, not 15-page reports you re-type
bar_raiser333 · reddit · 2026-11-21
A Reddit post shares prompting techniques to make deep research tools output ready-to-use deliverables instead of 15-page cited reports followed by an hour of manual data entry.
Key points:
- Start with the columns, not the question: specify exact fields (company, metric, value, period, source URL, quote) upfront, or pass a JSON schema.
- One claim and one source per row cuts fact-checking from an hour to minutes.
- Demand the period next to every number.
- Instruct "not found" instead of guessing.
- Don't run heavy mode on simple fact checks.
Includes a sample GLP-1 comparison prompt and recommends Valyu's DeepResearch, which returns .xlsx/docs/JSON matching the schema (somewhat promotional).
More from coding & agent
- Pydantic launches Agent Sandboxes in Logfire: millisecond cold starts, snapshot-resumable REPLs — samuelcolvin · 2026-11-21
- Building one customer-facing AI agent in-house costs $245K-$365K in year one — pritisinghhhh · 2026-11-21
- MEGA AI Engineering Course Wraps First Cohort with 2,000+ Developers Enrolled — johnlindquist · 2026-11-21
- OuterSSH open-sourced: SSH tunneling to run graphical agent apps from your phone — savvyRL · 2026-11-21
- You hit the kill switch but subagents keep running: Anaconda's agent incident tabletop — anacondainc · 2026-11-21
- Luke Wroblewski: why open an app to guide agents? Just use what's on screen — LukeW · 2026-11-21