Create Research Brief

Create Research Brief

Design multi-model research strategies for Claude Opus 4.6, Gemini 3.1 Pro Deep Research, and optionally GPT-5.2 Deep Research (site-restricted) or GPT-5.2 Chat across 9 research patterns.

Workflow

Phase 1: Research Design

1. Detect input type -- Direct request OR Research Request Specification (from research-interviewer)
2. Validate input -- Clarify if vague (skip if Specification present)
3. Detect model mode -- DUAL (default) / FULL / SINGLE
4. Detect OpenAI depth -- Deep Research (default for FULL) or Chat
5. Classify research pattern -- Use decision tree from references/pattern-registry.md to select one of 9 patterns
6. If Best Practices -- Load references/best-practices-dimensions.md and references/technology-profiles.md for 8-dimension framework
7. Assess topic risks -- Recency, contestation, source availability, false confidence, coverage gaps
8. Plan context budget -- Estimate combined output size, select context tier
9. Assign model roles -- Load Pattern x Model Configuration Matrix from references/model-profiles.md; assign roles, capabilities, thinking tiers, effort levels per pattern
10. Configure GPT-5.2 site restrictions -- If FULL mode, load site restriction library from references/model-profiles.md Section 6 and customize for topic
11. Configure Gemini file_search -- If user has relevant documents, load file search guidance from references/model-profiles.md Section 7 and recommend uploads
12. Design consolidation strategy -- Select pattern-default mode from Pattern Registry, allow user override
13. Generate prompts -- Use templates from references/model-profiles.md Section 5; for Best Practices, assemble dimension-specific fragments from references/best-practices-dimensions.md
14. Generate consolidation manifest -- Produce YAML manifest block per references/consolidation-manifest-schema.md
15. Produce brief -- Use output template from references/output-templates.md
16. Include merge prompt -- Append merge prompt referencing the manifest from references/output-templates.md

Phase 2: Consolidation (after user executes prompts)

17. Receive outputs -- User pastes/attaches research results
18. Consolidate -- Follow workflow in references/consolidation.md

Model Overview

Capability uniqueness: Claude = self-correction + cross-domain synthesis. Gemini = file_search. GPT-5.2 Deep = site-restricted search. Web research is a shared capability (Claude 84%, Gemini 85.9%); differentiate by approach, not access.

For detailed profiles, prompt templates, role assignment, and per-pattern configuration: references/model-profiles.md

Mode Detection

Model Mode

OpenAI Depth (FULL mode only)

GPT-5.2 Deep Research is the default for FULL mode. Site-restricted search is its unique differentiator -- use Chat only when explicitly downgraded.

Research Pattern Classification

Use the decision tree in references/pattern-registry.md to classify. Summary table:

Best Practices Pattern

When pattern = Best Practices, activate the 8-dimension framework:

1. Environmental Context -- Runtime, config, auth, secrets, project structure
2. Idiomatic Patterns -- How practitioners write it vs. how it merely works
3. Anti-Patterns & Guardrails -- What fails at scale, root cause reasoning
4. Testing & Validation -- Emulators, mocks, integration tests, gaps
5. Dependencies & Versions -- Compatibility matrices, packages to avoid
6. Operational Awareness -- Cost drivers, scaling, cold starts, monitoring
7. Decision Trees -- When to use which approach, architectural boundaries
8. Escape Hatches -- Known bugs, workarounds with expiration, when to eject

Weight dimensions by technology family from references/technology-profiles.md. Assemble prompts using dimension fragments from references/best-practices-dimensions.md.

Input Detection

Research Request Specification (from research-interviewer)

Detect by: YAML block with research_request: root containing objective:, questions:, scope:, constraints:.

When detected: Skip validation. Extract objective, questions.primary/secondary, scope.in_scope/out_of_scope, constraints.model_mode, constraints.depth_vs_breadth, context.prior_knowledge, metadata.suggested_research_type (verify, don't blindly accept).

Direct Request Validation

Required: Research objective specific enough to derive key questions.

Optional (with defaults): Research pattern (infer), context (none), timeline (standard), output use (general decision support), model mode (DUAL), OpenAI depth (Deep Research).

When insufficient and no Specification present:

To design an effective research strategy, I need:

1. Research objective: What specific question(s) do you want answered?
2. Research pattern (optional): Landscape / Comparative / Implementation / Best Practices / Competitive / Market / User / Economic / Compliance?
3. Context (optional): Background, constraints, or intended use?

Alternatively: Say "interview me about [topic]" to clarify your needs first.

Risk Assessment

Rate each factor High / Medium / Low:

Context Budget Planning

Output Estimates by Model

Combined Budget by Mode

Principle: Always prefer full unabridged outputs. Opus 4.6's 76% MRCR v2 long-context retrieval can attend throughout the window.

Consolidation Modes

Pattern-Default Mapping

For mode details and full consolidation workflow: references/consolidation.md

Effort & Thinking Directives

Include in Claude Opus 4.6 prompts:

# Claude API Configuration
model: "claude-opus-4-6"
thinking:
  type: "adaptive"
effort: "max"
max_tokens: 16000  # Up to 128000 for comprehensive output

If over-thinking detected: add "Use deep reasoning for strategic analysis; move efficiently through factual compilation."

Consolidation Manifest

Every research brief must include a consolidation manifest YAML block generated per references/consolidation-manifest-schema.md. The manifest travels from design through execution to consolidation.

Required fields: research_id, pattern, topic, objective, model_mode, created_at, models[] (per model: model_id, role, capabilities), pattern_metadata, coverage_matrix, consolidation (recommended_mode, pattern_default_mode, verification_priorities), research_chain, freshness (topic_volatility, confidence_half_life, staleness_indicators, recommended_refresh).

If this research follows prior research, populate research_chain.upstream_id with the previous manifest's research_id, research_chain.upstream_pattern, and research_chain.inherited_constraints.

Quality Gates

Before completing research brief:

• Input type detected (direct OR Specification)
• If Specification: fields extracted, interview_confidence noted
• If direct: objective is specific and actionable
• Model mode correctly detected (default: DUAL)
• Research pattern correctly classified (9 patterns, using decision tree)
• If Best Practices: 8 dimensions assessed, technology family identified, dimension weights applied
• Risk assessment completed (5 factors)
• Context budget estimated, tier selected
• Model roles assigned per Pattern x Model Configuration Matrix
• Claude prompt includes: effort directive, web search activation, cross-domain synthesis, self-review mandate
• Gemini prompt includes: thinking tier, structured data directives, citation requirements, table/matrix requests
• If FULL: GPT-5.2 site restrictions generated and customized for topic
• Gemini file_search recommendations included (if applicable)
• Consolidation mode matches pattern default (or documented override)
• Consolidation manifest generated with all required fields
• If sequential research: manifest includes research chain predecessors
• Freshness model populated (topic_volatility, confidence_half_life, staleness_indicators, recommended_refresh)
• Coverage matrix covers all key questions
• If from research-interviewer: prior_knowledge included in prompt context
• Merge prompt included at end of brief

Before completing consolidation:

• All outputs processed (full, unabridged if within budget)
• Claims triaged by type with provenance tags
• Disagreements resolved per protocol
• False confidence audit applied to unanimous high-risk claims
• Confidence tiers assigned (Tier 1 >75%, Tier 2 50-75%, Tier 3 <50%)
• Self-review pass completed
• Cross-domain synthesis section populated
• Gaps documented with recommended actions
• Effort level set to "max"
• Context tier documented in methodology

References