Question
Which prompt-engineering techniques from the Prompt Wiki should be added to a Claude Deep Research prompt investigating adoption, usage, sentiment, and competitive dynamics in AI-assisted development environments?
Answer
The original prompt already had strong behavioral constraints: broad discovery, no predetermined ranking, source traceability, evidence labels, contradiction handling, geographic-bias disclosure, and a defined report structure.
Four improvements materially strengthen it:
- Document the research process: Add an initial decomposition phase, a search plan, source notes by subquestion, a source-evaluation matrix, and a final issue-to-correction review table. This makes research quality auditable rather than merely asserted. This follows documented-research-process and gemini-deep-research-prompt.
- Separate search from synthesis: Explicitly separate broad discovery, source validation, critical synthesis, and report writing. This applies two-tier-search-architecture without requiring separate model calls; the separation is a process instruction that prevents the first narrative from controlling the final answer.
- Add a stopping rule: Stop when new credible sources no longer add material evidence, contradictions, or coverage of new regions or developer segments. Record why the search stopped and what remains under-researched. The documented-research pattern otherwise has a known weakness: it does not define convergence.
- Add a bounded self-review loop: Require a final pass that records
Issue found -> Correction made -> Remaining impact, without exposing private chain-of-thought. This applies reflexion to the report artifact rather than asking for hidden reasoning.
Two techniques should remain excluded:
- active-prompt and generate-knowledge-prompting can generate unsupported background knowledge before source validation. That conflicts with a research prompt whose priority is evidence grounding.
- gauntlet-loop-prompting is excessive for a single Claude Research run. Its worker-and-critic orchestration is slow and token-intensive; a bounded self-review is sufficient here.
The final prompt should therefore keep its existing research questions and evidence rules, while adding the four process controls above. It should also use constraint-based-prompting to state hard requirements and clarifying-questions only when missing information would materially change the research direction.
Related
- documented-research-process — makes research phases and quality artifacts visible
- two-tier-search-architecture — separates discovery and synthesis responsibilities
- constraint-based-prompting — converts quality expectations into explicit rules
- reflexion — supports bounded post-draft correction
- gemini-deep-research-prompt — concrete process-centric research prompt
- clarifying-questions — limits clarification before launch
Sources
^[raw/prompts/leaked/gemini-deep-research-prompt.md] ^[raw/prompts/leaked/perplexity-ai-system-prompt.md]