Documented Research Process

Definition (confidence: high)

A documented research process is a workflow pattern in which a research system prompt mandates an explicit, ordered set of research phases and requires the final report to contain the artifacts of those phases — making the process itself part of the deliverable. It is the “process over product” pattern: “The report is not just a final product; it’s a record of the research journey.”

Core Phases

The canonical example (the gemini-deep-research-prompt) defines five phases:

• TDP — Topic Deconstruction and Planning: decompose the question, list key concepts/ambiguities, pick relevant perspectives, justify sub-questions, and plan a search strategy with anticipated source-type biases.
• MIG — Multi-Faceted Information Gathering: diverse source search; per-sub-question Source Notes tables (Source / Date / Relevance / Key Findings / Methodology / Potential Biases / Conflicting Info).
• CAS — Critical Analysis and Synthesis: Source Evaluation Matrix (Credibility / Bias / Reliability), reasoned discrepancy analysis, integrative narrative with explicit citations, actionable gap identification.
• RG — Report Generation: mandatory report skeleton (Executive Summary → Introduction → Background → Subtopic sections → Impact → Recommendations → Conclusion → References) with the process tables embedded in subtopic sections.
• IR — Iterative Refinement: checklist-based self-review (derived from the prompt itself) driving a documented Gap/Issue → Action → Result table included in the final report.

Distinctive Features (confidence: high)

• Artifacts-as-output: the deliverable embeds source notes tables, evaluation matrices, and iteration logs — process transparency is structurally enforced.
• Show, don’t tell: the prompt repeatedly demands “show the reasoning, not just the conclusion.”
• Checklist-driven self-review: the IR phase checks each prior phase against criteria lifted directly from the prompt, closing the loop.
• No convergence criterion: iteration continues “as needed” without an explicit stopping rule — a known weakness.

Relationship to Other Patterns

• Contrasts with the invisible planning tier of two-tier-search-architecture: here the process is deliberately shown to the reader; there it is deliberately hidden.
• Complements reflexion: both iterate on self-critique, but reflexion corrects reasoning traces while the documented research process revises a whole report artifact.
• The phased structure is a task-state-machine: mandatory, ordered stages whose artifacts gate progress.

Implications

The documented research process is an output-structure strategy for enforcing quality: instead of gating research with tools (Anthropic’s anthropic-research-instructions tool-priority override) or hiding it (Perplexity’s two-tier architecture), it makes compliance verifiable in the text itself. The cost is verbosity and a heavier output-token budget; the benefit is an auditable evidence trail in every report.

Open Questions

• Does embedding mandatory tables meaningfully improve answer quality, or only perceived rigor?
• How should the pattern adapt when the user needs a short, direct answer rather than a full research dossier?
• Where does the pattern fail — e.g., topics with sparse sources, or time-sensitive queries where exhaustive review is too slow?

Sources

• raw/prompts/leaked/gemini-deep-research-prompt.md