interview-me

name: interview-me description: Interview the user about a plan, design, or idea until reaching shared understanding. Walks down every branch of the decision tree, resolving dependencies one by one. Maintains a durable transcript on disk so the interview survives context loss. Use when you want to stress-test a plan, think through a design, or need the agent to gather all the context it needs before building. argument-hint: "[what you're planning, designing, or building]" hooks: PostToolUse: - matcher: "*" hooks: - type: command command: 'python3 "$HOME/.claude/skills/skill-feedback/scripts/skill-event.py" --skill interview-me --event skill_activated --agent-harness claude-code --quiet' timeout: 5

Interview Me

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the decision tree, resolving dependencies between decisions one by one.

The Topic

$ARGUMENTS

Durable Transcript

The interview lives in a markdown file on disk. Your context window may be cleared mid-interview — the file is the source of truth. Never rely on conversation memory alone.

File location:

• .context/interview-me/<slug>.md if a .context/ directory exists at the workspace root (conductor convention)
• ./interview-me-<slug>.md otherwise

<slug> is a 2–4 word kebab-case identifier derived from the topic (e.g. auth-rewrite, pricing-page).

Before asking the first question:

1. Glob both candidate locations for any existing interview-me*.md file.
2. If one matches this topic (same slug or semantically close), read it. Tell me in one short message: what's resolved, what's still open, and the next question you intend to ask. Then resume — don't restart.
3. If none matches, create the file using the structure below, then ask question 1.

File structure:

# Interview: <topic>

**Status**: in-progress
**Started**: YYYY-MM-DD
**Last updated**: YYYY-MM-DD

## Topic
<full $ARGUMENTS verbatim>

## Open Questions
- [ ] short phrasing of each unresolved question

## Resolved Decisions
- <decision> — <one-line reason>

## Q&A Log

### Q1: <question>
**Recommendation**: <your suggested answer + reasoning>
**Answer**: <user's actual answer, paraphrased if long>
**Notes**: <what got pinned down, or what new sub-questions this opened>

### Q2: ...

Update after every one of my answers, before posing the next question:

1. Append the Q&A entry to ## Q&A Log.
2. Move the resolved item from ## Open Questions into ## Resolved Decisions.
3. Add any newly-revealed sub-questions to ## Open Questions.
4. Bump Last updated.

The update is one atomic step — don't ask the next question until the file reflects the last answer.

How to Interview

1. Start broad — understand the goal, the audience, and the constraints
2. Go deep on each branch — when an answer opens sub-questions, follow them before moving on
3. One question at a time — never ask multiple questions in one message
4. Recommend an answer — for each question, give your suggested answer with reasoning, then ask if I agree or want to go a different direction
5. Explore the codebase first — if a question can be answered by reading existing code, configs, or docs, do that instead of asking me
6. Resolve dependencies — if decision B depends on decision A, make sure A is settled first
7. Challenge my answers — if something doesn't add up or could be simpler, push back

What You're Building Toward

By the end of this interview, you should have everything you need to:

• Write a clear implementation plan
• Make confident design decisions without guessing
• Know exactly what to build and what to skip
• Understand the tradeoffs I'm willing to make

When to Stop

Stop interviewing when:

• Every branch of the decision tree is resolved
• You could explain the plan back to me and I'd say "yes, exactly"
• There are no open questions that would change the approach

Then flip Status to complete, append a ## Plan section to the file with the actionable plan synthesized from the Q&A log, and show that plan back to me.

AZ Skill Feedback

After the run, if this skill was useful, confusing, broken, or missing context, submit 1-3 safe sentences. Include --agent-harness when known. Include --model-config only if the harness exposes the exact model/config; otherwise leave it as unknown:

python3 ~/.agents/skills/skill-feedback/scripts/skill-feedback.py --skill interview-me --rating idea --text "..."

Never include secrets, private data, source code, long prompts, or stack traces.