create-research-brief

What this skill does

# 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](references/pattern-registry.md) to select one of 9 patterns
6. **If Best Practices** -- Load [references/best-practices-dimensions.md](references/best-practices-dimensions.md) and [references/technology-profiles.md](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](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](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](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](references/model-profiles.md) Section 5; for Best Practices, assemble dimension-specific fragments from [references/best-practices-dimensions.md](references/best-practices-dimensions.md)
14. **Generate consolidation manifest** -- Produce YAML manifest block per [references/consolidation-manifest-schema.md](references/consolidation-manifest-schema.md)
15. **Produce brief** -- Use output template from [references/output-templates.md](references/output-templates.md)
16. **Include merge prompt** -- Append merge prompt referencing the manifest from [references/output-templates.md](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](references/consolidation.md)

---

## Model Overview

| Model | Role | Key Capabilities |
|-------|------|------------------|
| **Claude Opus 4.6** | Primary Researcher | 84% BrowseComp, 68.8% ARC-AGI-2, 1M context, 128K output, self-correction, cross-domain synthesis |
| **Gemini 3.1 Pro Deep** | Structured Cataloger | 85.9% BrowseComp, 77.1% ARC-AGI-2, 1M context, 64K output, file_search (uploaded docs as sources), autonomous 5-30 min agent |
| **GPT-5.2 Deep Research** | Targeted Investigator | Site-restricted search (unique), leading MRCR v2 at 128-256K, 5-30 min autonomous agent, mid-session intervention |
| **GPT-5.2 Chat** | Recency Validator | Quick recent developments, community signals |

**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](references/model-profiles.md)

---

## Mode Detection

### Model Mode

| User Signal | Mode |
|-------------|------|
| No model specification | **DUAL** (default) |
| "all three models" / "full research" / "include OpenAI" / "comprehensive" | **FULL** |
| "without OpenAI" / "skip OpenAI" / "Claude and Gemini only" | **DUAL** (confirm) |
| "dual model" / "two model" | **DUAL** |
| "Claude only" / "quick research" / "fast analysis" | **SINGLE** |

### 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.

| User Signal | Depth |
|-------------|-------|
| No specification | **Deep Research** (default) |
| "quick OpenAI" / "lite" / "lightweight" / "skip deep research" | **Chat** |
| "OpenAI deep research" / "comprehensive OpenAI" / "site-restricted" | **Deep Research** (confirm) |

---

## Research Pattern Classification

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

| Pattern | Trigger Signals | Primary Deliverable | Default Consolidation |
|---------|----------------|--------------------|-----------------------|
| **Landscape Mapping** | "map the landscape", "who's out there", "ecosystem overview" | Taxonomy + player inventory + white space map | breadth_first |
| **Comparative Evaluation** | "compare X vs Y", "which should I choose", "trade-offs" | Weighted decision matrix + sensitivity analysis | confidence_weighted |
| **Implementation Pattern** | "how do I implement", "architecture patterns", "reference architecture" | Architecture decision catalog + pattern catalog + anti-pattern register | depth_first |
| **Best Practices** | "best practices for", "idiomatic patterns", "gotchas", "anti-patterns" | 8-dimension technology knowledge base | gap_driven |
| **Competitive Intelligence** | "competitive moat", "positioning analysis", "competitive dynamics" | Per-competitor strategic profile + dynamics analysis | adversarial |
| **Market Research** | "market size", "TAM", "segments", "entry strategy" | TAM/SAM/SOM + segmentation framework + dynamics | standard |
| **User Research** | "user needs", "JTBD", "persona", "pain points" | Persona profiles + JTBD map + ranked unmet needs | depth_first |
| **Economic Analysis** | "ROI", "TCO", "cost-benefit", "business case" | Financial model + sensitivity analysis + benchmarks | confidence_weighted |
| **Compliance & Requirements** | "compliance", "regulatory", "GDPR", "audit requirements" | Requirements register + constraint map + governance | gap_driven |

### 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](references/technology-profiles.md). Assemble prompts using dimension fragments from [references/best-practices-dimensions.md](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