Gemini Deep Research Prompt
This capture presents a shared system prompt for a Gemini Deep Research assistant, circulated as a public gist by GitHub user swayson. Unlike the leaked Anthropic and xAI prompts in this wiki, it comes from a third-party gist rather than a known leaks mirror, so provenance is unverified — confidence: low.
The prompt is notable for being tool-agnostic and process-centric: it contains no tool schema or XML envelope. It governs the reasoning and output structure of a research assistant, requiring every report to make the research process itself visible to the reader.
Process Overview: TDP / MIG / CAS / RG / IR
The prompt mandates five documented phases that every report must explicitly reflect:
- Topic Deconstruction and Planning (TDP) — decompose the question, list key concepts and ambiguities, select relevant perspectives (historical, economic, ethical, etc.), justify sub-questions, and plan a search strategy including anticipated source-type biases.
- Multi-Faceted Information Gathering (MIG) — diversify sources and record per-sub-question Source Notes tables (Source, Date, Relevance, Key Findings, Methodology, Potential Biases, Conflicting Info).
- Critical Analysis and Synthesis (CAS) — produce a Source Evaluation Matrix (Credibility / Bias Level / Overall Reliability), analyze discrepancies (why they arose, not just that they exist), synthesize narratively with explicit citations, and identify actionable gaps.
- Report Generation (RG) — enforce a mandatory structure (Executive Summary → Introduction → Background → Subtopic sections → Impact sections → Recommendations → Conclusion → References) with the process tables embedded inside the subtopic sections.
- Iterative Refinement (IR) — review the draft against a checklist derived from the prompt itself, then document changes in a Gap/Issue → Action → Result table included in the final report.
Implications
The prompt converts a research workflow into a task state machine: phases are mandatory, ordered, and their artifacts are required outputs. This makes “the report is the process” literal — process compliance is enforced through output structure rather than through tool gating.
Transparency as Deliverable
Every phase produces a visible artifact inside the report: breakdowns are shown not summarized, source notes are tabulated, credibility is rated with justification, and discrepancies are reasoned through. The prompt repeatedly instructs the model to “show the reasoning, not just the conclusion.”
Implications
This is the opposite of Perplexity’s two-tier-search-architecture, where the planning tier’s work is deliberately invisible to the reader. Here the process is the product — a design choice suited to an audience that wants to audit research quality.
Bias Mitigation in Three Stages
Bias awareness is required three separate times: (1) preemptively in TDP, listing potential biases per anticipated source type; (2) per-source in the MIG notes table; (3) rated again in the CAS evaluation matrix. The prompt treats bias identification as an explicit, checkable step rather than a post-hoc filter.
Implications
The three-stage design is an instruction-level analog of a quality gate: it forces the model to state its epistemic assumptions before, during, and after gathering evidence.
Style and Formatting Rules
The prompt mandates responding in the user’s language, a formal/objective/academic tone, Markdown formatting, a consistent citation style, bold for key words, and date awareness (current date at top; publication dates used to judge timeliness). These are pure presentation rules with no tool dependence — the same contract could ride on any model.
Open Questions
- Whether Google’s actual Gemini Deep Research ships this exact prompt — provenance is a third-party gist, unauthenticated.
- Whether the heavy process overhead (mandatory tables, documented iterations) is acceptable for small research tasks, or if the prompt would be better gated to large topics.
- How the iterative-refinement loop terminates — the prompt says “repeat as needed” but sets no explicit convergence criterion.
Relationships
- “references” anthropic-research-instructions — sibling research-mode prompt (Anthropic); contrast in tool-gating vs process-gating
- “differs from” two-tier-search-architecture — visible process vs invisible planning tier
- “implements” documented-research-process — the process-over-product workflow pattern distilled here
- “belongs-to” gemini — Gemini model family context
- “uses” markdown-header-format — Markdown heading structure for the report
- “references” reflexion — iterative self-review loop, applied to the whole report rather than reasoning steps
Sources
- raw/prompts/leaked/gemini-deep-research-prompt.md