|
The Opinionated Daily of Technology & Code
GARRY'S TAKE
|
| Tuesday, March 3, 2026 · Decision Memo |
DECISION MEMO |
|
ANALYSIS
Prompt Engineering for Deep Research
Three approaches to getting high-signal reports from Deep Research tools, compared.
By Garry Sandcastle
|
|
Deep Research follows a "you state the outcome, you control sources, it proposes a plan, it executes, you receive a structured report with citations" loop. Your prompt is less like asking a question and more like commissioning a scoped research deliverable. Source choice is the highest-leverage prompt engineering move. |
|
BEST PICK
Narrow + Deliverable-Specific Prompts
When you can list the top 5 sub-questions yourself, go narrow and precise. Specify the decision/use-case ("I'm using this to decide X"), scope boundaries (geography, time window, what's out of scope), deliverable constraints (required sections, comparisons), and evidence standards. This reduces time wasted on irrelevant branches and improves citation density.
|
ALTERNATIVE
Broad + Phase-Separated Prompts
When you cannot list the sub-questions, start broader but request a "map first" output. Ask for landscape/overview first, then deep dives. Deep Research supports plan review and modification before running, and you can interrupt mid-run to refine focus.
|
ALTERNATIVE
Iterative Refinement via Multi-Turn
Use an initial broad pass, review the plan output, then refine in follow-up turns. Works well when the research domain is unfamiliar. Trade-off: higher token cost and longer wall-clock time, but often surfaces unexpected angles a narrow prompt would miss.
|
|
TRADEOFF COMPARISON |
| Approach | Speed | Depth | Citation Density | Best When |
|---|
| Narrow + Specific | Fast | Deep | High | Known domain | | Broad + Phased | Medium | Wide then deep | Medium | Unknown territory | | Iterative Multi-Turn | Slow | Variable | Medium-High | Complex/unfamiliar |
|
|
THE VERDICT
For operators who already know their domain, narrow prompts with explicit deliverable structure win on both speed and quality. Use broad prompts only when genuinely exploring unfamiliar territory -- and always request a plan before the full run.
|
You are receiving this because you subscribed to Garry's Take.
If this issue was forwarded to you, subscribe for direct delivery. |