Verified against Perplexity Pro · 2026-07-22
Turn a vague research question into a Deep Research report with a real evidence trail
A brief for Perplexity Deep Research mode that forces a stated research plan, named source priorities, a recency window, and a Finding/Source/Date/Confidence table instead of a fluent-sounding single-pass answer.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Run this as a Deep Research task, not a quick search. Before pulling any sources, state the research plan you intend to follow — the sub-questions you will answer and the search angle for each one — in three to five lines, then proceed with that plan. If you have to revise the plan mid-run because early results point somewhere the original plan didn't anticipate, say so explicitly rather than quietly following the new direction without flagging the change.
RESEARCH QUESTION
Is on-device small-model inference actually cheaper than cloud API calls at production scale for a mid-size SaaS company in 2026?
WHAT THIS IS FOR
Deciding whether to migrate our support-ticket classifier off a hosted API onto local inference this quarter.
SOURCE REQUIREMENTS
- Prioritize: independent benchmark writeups and infrastructure cost breakdowns, not vendor marketing pages
- Actively deprioritize: affiliate-linked "best AI tools" listicles and any article that reads as sponsored content — these may appear in search results but should not anchor a finding on their own.
- Recency window: only weight sources inside last 9 months as current evidence. Older sources may be cited for background or historical context, but must be explicitly labeled as background, not as support for a current claim.
- Do not treat one outlet's framing of a fact as the fact itself if other outlets, or the underlying primary data, frame it differently — surface the disagreement in the framing rather than silently adopting the first version you found.
WHAT WOULD CHANGE THE ANSWER
Before finalizing, note explicitly what kind of new evidence — a specific study, a specific data release, a specific event — would be strong enough to change the answer you're about to give. If nothing you found would plausibly change your answer, say that too; it's a useful signal about how settled the question actually is.
FINAL REPORT FORMAT
1. Answer at a glance — three to five sentences, with no hedge word ("may," "could," "possibly") that isn't backed by an actual, named disagreement in the sources rather than generic caution.
2. Evidence table — one row per key finding: Finding | Source | Date | Confidence (High/Medium/Low). Confidence should reflect how many independent sources agree and how directly each one addresses the specific finding, not how confidently any single source states it.
3. Where the evidence is thin or contested — name the specific gap (a missing data point, a methodology disagreement, a claim only one outlet makes) rather than a generic "more research is needed."
4. Full source list in the order first cited, with publication dates.
If the original question could reasonably be scoped two different ways — a narrower reading and a broader one — note the alternate scoping as an aside and explain which one you answered, rather than silently picking one and hiding that a choice was made.Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Perplexity Deep Research runs an iterative loop of dozens of searches and page reads, forming and revising a plan mid-run before it writes the final synthesis — asking it to state that plan up front, in the sub-questions and search angles it will chase, anchors what the planning step actually searches for instead of letting it default to whatever the first page of results happens to convert into steps. Deep Research trades latency for exhaustiveness by design, so naming a decision context and a recency window is a budget instruction, not decoration: it tells the tool what to spend that extra time on rather than spreading equally thin coverage across an unbounded question, and it tells the tool which of two technically-correct-but-differently-scoped answers is the one you actually need. Forcing a Finding/Source/Date/Confidence table instead of prose also surfaces disagreement structurally — two rows with the same Finding column and different Confidence values are visibly in tension, where the same disagreement buried in a paragraph reads as one smooth, and misleadingly certain, narrative. The 'what would change the answer' step does something a report format alone cannot: it forces the model to commit to a falsifiability condition instead of hedging indefinitely, which is the actual difference between a research report that helps a decision get made and one that reads well but leaves every option still open. And the deprioritized-source-types field matters because Deep Research's retrieval step will happily surface a well-optimized listicle or a sponsored comparison page if it ranks well for the query — naming what to discount doesn't remove those pages from the search results, but it stops them from anchoring a Confidence rating they haven't earned.
What you get back
Plan: (1) compare per-token cost of hosted API vs local GPU amortized cost at our volume, (2) check independent benchmarks for accuracy parity on classification tasks, (3) check operational overhead reports from teams who migrated. Answer at a glance: At sub-1M requests/month, hosted APIs remain cheaper once you include GPU idle time and maintenance; the crossover point independent benchmarks report is closer to 5-10M requests/month for a classification-sized model. Evidence table (excerpt): "Crossover near 5-8M req/mo for 7B-class models" | independent infra benchmark, June 2026 | High confidence. What would change this: a sustained drop in GPU spot pricing below current levels, or a hosted provider cutting per-token pricing by more than 30% — neither appears imminent based on current vendor announcements.
Verified against
Perplexity Pro Deep Research (Sonar-based) · 2026-07-22
Changelog
- 2026-07-22 — Initial publish, verified against Perplexity Pro Deep Research mode.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
