# Inside Perplexity’s Deep Research System Prompt: A Full Breakdown
Deep Research has become one of the most talked-about capabilities in the AI space. The idea—an assistant that reads dozens of sources, weighs the evidence, and produces a genuinely thorough report instead of a quick answer—has quickly moved from novelty to expectation. Perplexity was among the earliest products to ship a production-grade Deep Research feature, and the system prompt that governs its behavior offers a rare, concrete look at how a leading AI search company orchestrates a long-form generation pipeline.
System prompts for top-tier products are hard to come by. Companies treat them as competitive IP, and the prompts rarely surface in clean, readable form. When one does, it becomes a valuable reference for anyone building AI applications—whether you are designing an assistant, planning product features, or refining your own prompt strategy. The practices of established players like Perplexity, Microsoft, and Google set a useful benchmark, and studying them is one of the fastest ways to level up your own work.
This article walks through the Perplexity Deep Research system prompt (dated April 23, 2025), section by section, and translates its design decisions into practical takeaways for developers and product teams.
## What the Prompt Tells the Model to Be
The prompt opens with a tightly scoped role definition. The model is told: *”You are Perplexity, a helpful deep research assistant trained by Perplexity AI.”* That single sentence does a lot of work. It fixes the assistant’s identity, its employer, its disposition (helpful), and its specialty (deep research) before any task instruction appears.
Immediately after, the goal block sets the deliverable: in response to a user query, the assistant must produce a **long, comprehensive, well-structured research report**. Three modifiers are stacked deliberately—long, comprehensive, well-structured—and they reappear as enforceable constraints throughout the prompt. The target audience is specified as academic, and the model is instructed to prioritize verbosity so that no relevant subtopic is overlooked. A hard floor of **10,000 words** is set as the minimum output length.
The takeaway for prompt designers is that Perplexity front-loads identity and deliverable expectations. The model never has to infer what it is or what a successful output looks like—the definition is locked in before any procedural instructions begin.
## The Report Format and Document Structure Rules
The `
### Mandatory Markdown Hierarchy
Every report must begin with a single `#` title, organize major themes into `##` sections, break those into `###` subsections, and reserve `####` for special cases only. The prompt explicitly forbids skipping header levels—an easy mistake when models auto-generate structure. Beyond the hierarchy, the prompt demands at least five `##` body sections and requires a narrative paragraph inside every section before the next one begins.
### The Ban on Lists
One of the most striking choices is the blanket prohibition on bullet points and lists. The prompt repeats this instruction across multiple sections: *”Never use lists, instead always use text or tables.”* Comparative data should appear in tables; everything else must be rendered as flowing prose. The rationale is that lists break narrative flow and let the model offload thinking into disconnected fragments. Forcing prose ensures that each point is connected to the next with reasoning, not just adjacency.
### Paragraph Depth Requirements
Each paragraph is required to contain **at least four to five sentences**, present novel insights grounded in source material, connect back to the original query, and build on previous paragraphs to create a continuous narrative. This is a quality control mechanism: it pushes the model away from shallow, list-like paragraph stubs and toward genuine analysis. The mandatory section flow also requires an opening summary paragraph before the main report and a conclusion section that synthesizes findings and suggests next steps.
For developers, the lesson is that structure is not an afterthought. Perplexity treats it as a first-class constraint, repeated and reinforced, because format failure is one of the most visible ways a long-form output can degrade.
## Citation Rules and Source Handling
The `
Importantly, the prompt forbids appending a References section or sources list at the end. The rationale: Perplexity’s UI already renders the source list to the user, so a duplicate list in the report body would be redundant. The model is also told not to reproduce copyrighted material verbatim and to fall back on existing knowledge when search results are empty or unhelpful.
This citation architecture is instructive. It enforces **inline, sentence-level provenance** rather than endnotes, which keeps every claim tethered to a source right where it appears. For teams building retrieval-augmented generation (RAG) products, this is a strong template: citations should be structural, not decorative.
## Planning, Personalization, and Guardrails
The `
The `
The `
Finally, the `
## What Developers and Product Teams Can Learn
Studying this prompt yields several transferable lessons for anyone building AI products:
– **Lock identity early.** A precise role definition up front prevents identity drift over long generations. State who the assistant is, who trained it, and what its specialty is before any procedural rules.
– **Make structure a constraint, not a suggestion.** Perplexity enforces header hierarchy, minimum section counts, and paragraph depth as hard rules. Vague instructions like “write a detailed report” produce inconsistent output; explicit format rules produce reliable structure.
– **Ban the shortcuts you don’t want.** The prohibition on lists is repeated across four separate sections. When a failure mode is common, redundancy is a feature, not a bug.
– **Wire citations into the structure.** Sentence-level inline citations with bracketed indices keep provenance visible and verifiable. If your product surfaces sources, design the prompt so citations are structural, not optional.
– **Protect privacy and IP in the prompt.** Explicit instructions against revealing the system prompt, disclosing personalization details, and reproducing copyrighted material verbatim are guardrails that belong in the prompt itself, not just in a content filter layered on top.
– **Inject runtime context.** A hardcoded current date and timestamp anchors the model to the present. Any temporal-sensitive product should do the same.
– **Force a planning phase.** Requiring the model to break the task into steps and verbalize its plan before writing improves coherence and gives users visibility into the reasoning process.
## A Note on Recency and Accuracy
This system prompt was captured on April 23, 2025. AI companies iterate on their prompts frequently, and the version governing Perplexity’s Deep Research today may differ substantially—sections may have been added, tightened, or removed entirely. Treat this breakdown as a snapshot of one point in time, not a live specification. If you are building a product and want to benchmark against current behavior, test against the live product rather than relying on any single leaked prompt. System prompts are moving targets by design.
## Conclusion
The Perplexity Deep Research system prompt is a masterclass in constraint-driven prompt design. Every section—from the opening role definition to the final output directive—reinforces a small set of principles: be comprehensive, be structured, cite everything, and never fall back on shortcuts. For developers and product managers working on AI applications, it is a concrete reminder that the gap between an average assistant and a great one often comes down to how rigorously the system prompt is engineered. The prompt is not just an instruction to the model; it is a product spec written in natural language, and studying the best examples is one of the most efficient ways to improve your own.
