Interviewing Plans

Overview

Plans fail because of hidden assumptions, not missing features. This skill systematically interviews stakeholders to surface ambiguities before they become bugs.

Core principle: Every vague term is a bug waiting to happen. "Make it faster" means nothing until someone says "reduce P95 latency from 800ms to 200ms."

When to Use

Trigger conditions:

• User invokes /interview-plan <file>
• Plan contains undefined qualitative terms ("better", "faster", "modern", "relevant")
• Requirements reference external artifacts without links ("design has mockups")
• Technical approach lacks specifics ("add caching", "use React")
• Timeline exists without milestone breakdown
• User complaints mentioned without specific pain points

Do NOT use for:

• Already-detailed specs with concrete acceptance criteria
• Simple bug fixes with clear reproduction steps
• Tasks where user explicitly wants implementation, not planning

Required Categories

Interview is NOT complete until ALL categories are covered:

Category	Must Surface
Technical	Architecture decisions, data flow, state management, API contracts
UX/Design	User flows, states (loading/empty/error), mobile strategy, accessibility
Tradeoffs	What are we NOT doing? What's MVP vs future? What's negotiable?
Edge Cases	High-volume users, empty states, failure modes, concurrent access
Security	Auth/authz, data exposure, input validation, audit requirements
Testing	How do we know it works? Unit/integration/E2E split, test data
Rollback	What if it fails in prod? Feature flags? Parallel running?
Dependencies	Who owns what? What's blocking? Who approves?

Interview Protocol

1. Read and Analyze

Read the plan file. For each section, identify:

• Undefined terms: Words that mean different things to different people
• Missing specifics: Numbers, formats, thresholds not specified
• Assumed context: References to things not in the document
• Hidden dependencies: Things that must be true for the plan to work

2. Ask in Batches by Topic

Use AskUserQuestion with 2-4 related questions per batch. Group by category.

Before each batch, state your reasoning:

I notice the plan says "[quote]" but doesn't specify [gap].
This matters because [consequence if undefined].

3. Challenge Vague Answers

When user gives a deflecting or vague answer:

User Says	You Respond
"We'll figure it out later"	"What specifically needs to be true before we can figure it out? Who decides?"
"It should be fast enough"	"What's the current latency? What would make you say 'this is too slow'?"
"Standard approach"	"Which standard? Can you point to an example in this codebase?"
"Design will handle that"	"Has design delivered this? If not, when? What's our fallback?"
"Users will understand"	"Which users? Have we tested this assumption? What's the support cost if wrong?"
"I trust you to figure it out"	"My job is to surface gaps, not make assumptions. What would happen if I guess wrong on [specific item]?"
"We're running out of time"	"I can compress into 2 more batches. But skipping categories means shipping bugs. Which risks do you want documented as 'owner: you'?"
"Just start building"	"Building with gaps means rework. 5 more minutes now saves days later. What's the real constraint - time or energy?"

Never accept:

• Qualitative terms without quantitative definitions
• Future tense without owners and dates
• References to artifacts without links or locations
• "Someone" instead of a specific name/team

4. Probe for Non-Obvious Gaps

Obvious questions (avoid these as primary questions):

• "What are the requirements?"
• "What's the timeline?"
• "Who's the user?"

Non-obvious questions (these surface real gaps):

• "What happens to in-flight requests during deployment?"
• "If this succeeds beyond expectations, what breaks first?"
• "What's the most expensive assumption we're making?"
• "Which part of this plan are you least confident about?"
• "What would make you cancel this project tomorrow?"
• "Who will be upset when this ships, and why?"
• "What's the first thing users will try that we haven't designed for?"

5. Track Coverage

Maintain mental checklist. After each batch, assess:

• Technical architecture clear?
• UX states defined?
• Tradeoffs explicit?
• Edge cases identified?
• Security considered?
• Testing strategy exists?
• Rollback plan defined?
• Dependencies confirmed?

Handling Time Pressure

When stakeholder is impatient:

Compress, don't skip. Combine remaining categories into 1-2 dense batches. Never skip a category entirely.

Make risk explicit. If they insist on skipping: "I'll document [Security: TBD, Owner: @you] in the spec. You're accepting this risk."

Name the tradeoff. "5 more minutes now, or 2 days of rework when we discover [gap] in production. Your call."

Never rationalize:

• "They know their business" - They don't know what they haven't thought about
• "I can figure it out" - Your assumptions are bugs waiting to happen
• "They seem confident" - Confidence is not evidence of completeness
• "I'm being annoying" - Surfacing gaps IS the job

Completion Criteria

Interview is complete when:

1. All vague terms operationalized - Every qualitative word has a number or concrete definition
2. All categories covered - Checklist above is complete
3. No deflections remaining - All "we'll figure it out" answers have been challenged and resolved
4. Dependencies confirmed - External people/teams/artifacts are named and accessible
5. You cannot find more gaps - Genuinely exhausted questioning, not just tired of asking

Output Format

When interview is complete, write structured spec:

# [Project Name] - Implementation Spec

## Overview
[1-2 sentences: what this is and why it matters]

## Success Criteria
- [ ] [Measurable criterion with number]
- [ ] [Measurable criterion with number]

## Technical Decisions
| Decision | Choice | Rationale |
|----------|--------|-----------|
| [Area] | [What we're doing] | [Why, including rejected alternatives] |

## User Experience
### Happy Path
[Step by step flow]

### Edge Cases
| Scenario | Behavior |
|----------|----------|
| [Case] | [What happens] |

### Error States
| Error | User Sees | Recovery |
|-------|-----------|----------|
| [Type] | [Message/UI] | [How to fix] |

## Security Considerations
- [Item with mitigation]

## Testing Strategy
- **Unit**: [What's covered]
- **Integration**: [What's covered]
- **E2E**: [Critical paths]

## Rollback Plan
[How we undo this if it fails]

## Open Questions
- [ ] [Anything still unresolved, with owner]

## Dependencies
| Dependency | Owner | Status |
|------------|-------|--------|
| [Thing] | [Person/team] | [Confirmed/Pending/Blocked] |

Red Flags - Dig Deeper

When you hear these, the interview is NOT done:

• "That's a good question, we should think about that"
• "I'm not sure, but probably..."
• "The other team handles that"
• "We've always done it this way"
• "That won't happen" (about edge cases)
• "We'll add that later"
• "It's just like [other project]" (without specifics)

Common Mistakes

Mistake	Fix
Accepting first answer	Always ask "why?" at least once
Asking obvious questions	Lead with non-obvious probes
Moving on when stuck	Challenge deflections explicitly
Single questions	Batch 2-4 by topic
Not showing reasoning	Explain why each question matters
Stopping when tired	Check completion criteria, not energy level