/interview

Structured conversational extraction of process knowledge from a subject matter expert. Produces a MECE decomposition through guided dialogue.

Usage

/interview <process or workflow to extract>

Examples:

• /interview our invoice approval workflow
• /interview the deployment process for our SaaS platform
• /interview then describe what process you want to map

How It Works

STANDALONE (always works)
- Structured 5-phase interview protocol
- Adapts to your communication style
- Produces MECE tree + Agent SDK JSON

SUPERCHARGED (when MCP server connected)
+ Interactive tree visualization as you build it
+ Click nodes to refine, validate scores live

Interview Protocol

Phase 1: Context and Scope (3-5 questions)

Establish what we are decomposing:

• What is the process and who is involved?
• What triggers it and when is it done?
• What is explicitly excluded?

Phase 2: Happy Path Walkthrough (open-ended)

You narrate the end-to-end process as it works when everything goes right. I probe each step for:

• Who does it (actor)
• How long it takes (duration)
• What goes in and what comes out (inputs/outputs)
• Is it conditional or always happens (conditionality)
• Can things happen in parallel (parallelism)

Phase 3: Exception Discovery (per step)

Go back through each step and surface:

• What fails and how
• Edge cases and skip conditions
• Branching logic

Phase 4: Boundary Conditions

Discover handoffs, parallel activities, upstream dependencies, downstream consumers, time constraints, and approval gates.

Phase 5: Validation

Present the decomposition tree and iterate until you confirm it covers the process completely without overlap.

Adaptive Behavior

I adjust interview style based on how you think:

• Structured thinker: I match your top-down approach, validate MECE against your mental model
• Narrative thinker: I let you tell stories, organize afterward, use your stories as validation scenarios
• Time-constrained: I show a draft decomposition and let you correct it (Phase 5 first)

Output

Same dual output as /decompose:

1. Human-readable tree (markdown)
2. Agent SDK JSON
3. Interactive visualization (when MCP server connected)

Tips

1. Don't worry about structure -- just describe how things work. I'll organize it.
2. Include the messy parts -- exceptions and edge cases are the most valuable.
3. Say "it depends" -- that's a branching point, and I'll probe for conditions.
4. Correct me freely -- the tree is wrong until you say it's right.