/workflow-research

name: workflow-research
description: "Research and plan a knowledge work task thoroughly before execution. Use when starting any significant piece of work - drafting important communications, making strategic decisions, preparing for meetings, or tackling analysis projects. Triggers on requests like 'help me prepare for', 'I need to draft', 'plan out', or any high-stakes knowledge work."

Research: Knowledge Work Research Workflow

You are orchestrating the **Research phase** of the Compound Knowledge Work loop. Your job is to thoroughly research a task and transform it into an actionable plan that makes execution straightforward.

**80% of compound knowledge work is in research and review.** Do not rush this phase.

Process

Phase 0: Discover Available Tools

Before starting research, check what MCP tools and CLIs are available in the environment (e.g., email, calendar, meeting notes, CRM, company intel). Note which are available and pass this to all agents so they can use them proactively.

Phase 1: Understand the Task

1. **Read the task description carefully.** If the user provided a document, file, or URL, read it thoroughly. 2. **Classify the work type:**

• Communication (email, memo, announcement, presentation)
• Decision (strategic choice, vendor selection, resource allocation)
• Analysis (market research, financial analysis, competitive review)
• Meeting (preparation, facilitation, follow-up)
• Coaching (leadership challenge, personal effectiveness)
• Operations (process design, project planning, documentation)

3. **Determine stakes and complexity:**

• **Low stakes**: Internal notes, routine updates -> lighter planning
• **Medium stakes**: Team communications, standard decisions -> moderate planning
• **High stakes**: Board communications, strategic decisions, sensitive topics -> full planning

Phase 1.5: Ask Clarifying Questions

Before launching research, identify gaps and ask the user clarifying questions. This prevents wasted research on the wrong problem.

**When to ask questions:**

• Task description is vague or has multiple interpretations
• Success criteria are unclear
• Stakeholders or constraints aren't specified
• High-stakes work where misunderstanding would be costly
• User's stated request might differ from their actual need

**Question patterns (from superpowers):**

• **Ask one at a time**, wait for answer before next question
• **Include your hypothesis**: "What does success look like? My hypothesis: [X]"
• **Focus on research-direction questions**: What would change your approach?
• **Catch red flags early**: Political sensitivity, irreversibility, high stakes

**Hard Gate for high-stakes work** - You MUST get answers to: 1. Who is the audience/stakeholder? 2. What does success look like? 3. What constraints exist (political, budget, timing)? 4. What have you tried before that didn't work?

**Example questions:**

• "What does success look like for this communication? My hypothesis: Board approval vs. team buy-in require different approaches"
• "Who is the decision-maker here? This affects stakeholder mapping"
• "What's the timeline? This determines research depth"
• "Are there political sensitivities I should know about?"
• "What have you already tried? What didn't work?"

**When to skip this phase:**

• Low-stakes routine work (internal notes, simple updates)
• User provided comprehensive brief with explicit success criteria
• Task is crystal clear and unambiguous

Phase 2: Search Past Learnings

Before any new research, check what we already know. This is where compounding pays off.

**In Claude Code** (file-based retrieval): 1. **Check the learnings index** at `.context/learnings/INDEX.md` - scan for matching category, type, and tags. 2. **Search by category**: Grep `.context/learnings/[category]/` for the relevant work type (e.g., `communication/`, `decision/`, `meeting/`). 3. **Search by type**: Look specifically for `pattern` and `template` insights that match the current task. 4. **Search by keywords**: Grep across all learnings for topic-specific terms (stakeholder names, project names, framework names). 5. **Search for preferences**: Grep for `type: preference` insights in the relevant category - these capture style, tone, and detail preferences from previous work.

If past learnings exist, incorporate them into the plan:

• **Reuse patterns** (type: `pattern`) that worked before
• **Avoid failures** (type: `failure`) that were documented
• **Apply templates** (type: `template`) if a relevant one exists
• **Honor preferences** (type: `preference`) from previous work in this category

Surface what you found: "Based on previous [category] work, we learned [X] and have a template for [Y]."

Phase 3: Parallel Research

Launch research agents in parallel based on the work type. Use the Task tool with appropriate subagents. **Pass the stakes level to each agent** so they can calibrate depth.

**Scale agents to stakes:**

| Stakes | Agents to Run | |--------|---------------| | **Low** | context-gatherer only (brief scan, no web search) | | **Medium** | context-gatherer + 1-2 agents from the work type table below | | **High** | Full agent roster from the work type table below |

**Agent roster by work type** (for medium/high stakes):

| Work Type | Additional Agents | |-----------|-------------------| | Communication | stakeholder-mapper, precedent-researcher | | Decision | stakeholder-mapper, constraint-analyzer, precedent-researcher | | Analysis | constraint-analyzer, precedent-researcher | | Meeting | stakeholder-mapper, precedent-researcher | | Coaching | stakeholder-mapper | | Operations | constraint-analyzer, precedent-researcher |

**For high-stakes work, always add:**

• External research via web search for best practices, frameworks, and examples

Phase 4: Research Decision

After parallel research returns, assess whether you need additional external research.

**Always research externally w