Interview

Interview

Run a focused discovery interview before creating a PRD, writing a plan, shaping UX, or starting implementation. Ground in the repo first, then run grilling for the decisions that cannot be inferred.

This skill does not write code, create issues, or produce a final PRD by default. It produces an interview brief. Recommend the next skill; do not invoke another user-invoked skill.

Contract

Inputs:

• Rough feature idea, issue number, product decision, bug class, or architecture question.
• Optional docs, links, transcripts, screenshots, or existing tracker context.

Outputs:

• Concise context scan summary.
• Settled decisions from grilling.
• Final interview brief ready for prd-writer, feature-intake, shape, spec-first, or direct implementation.

Creates/Modifies:

• None by default.
• May write tracker comments, PRD bodies, or memory files only when explicitly requested after the interview.

External Side Effects:

• None by default.
• Reads local repo context and, when needed, tracker or linked documentation.
• Writes external systems only after explicit approval.

Confirmation Required:

• Before creating or editing GitHub issues, PRDs, comments, memory files, or other durable artifacts.

Delegates To:

• grilling for the design-tree interview (frontier rounds, recommended answers).
• domain-modeling when a term crystallizes or conflicts with CONTEXT.md.

Recommend next (do not invoke): prd-writer, feature-intake, shape, spec-first, prd-quality-gate.

When To Use

• A user asks for /interview, "grill me", "grill me with docs", or equivalent.
• A feature idea is too vague to turn directly into a PRD.
• Existing repo docs probably answer part of the question, but missing decisions still need the user.

Skip this skill when:

• The user already provided a complete PRD or issue with acceptance criteria.
• The request is a small, obvious edit and the user said to implement directly.
• The only missing context is discoverable from the repo with no user decision.

Workflow

1. Ground In Repo Context

Read repo context before asking questions:

• Start with .agents/README.md when present.
• Read relevant .agents/memory/ files, especially .agents/memory/memory.md, .agents/memory/context.md, and any task-relevant .agents/memory/system/ docs.
• Read CONTEXT.md / CONTEXT-MAP.md and docs/agents/domain.md when present.
• Check recent .agents/sessions/ entries only when they are relevant to the topic.
• Read the applicable AGENTS.override.md / AGENTS.md chain for routing and repo rules. Read CLAUDE.md when the active workflow is Claude-specific.
• Search docs, README files, source code, and issues for the topic before asking the user to repeat known context.

Do not look for a local plans directory under .agents; plans live on GitHub issues and PR comments.

When the user provides external docs or says "with docs", read only the relevant sections and keep a short source list for the final brief.

2. State What Is Known

Before asking questions, summarize the context scan in three compact bullets:

• What the repo already says.
• What is still ambiguous.
• Which downstream artifact this interview is likely feeding.

If the repo gives enough context, ask for confirmation instead of running a long interview.

3. Run grilling

Run the grilling skill on the remaining decisions. It owns the design tree, the frontier, recommended answers, and the facts-vs-decisions split.

When a term is resolved or conflicts with CONTEXT.md, run domain-modeling inline.

4. Stop At The Right Time

Stop when one of these is true:

• The grilling frontier is empty and the brief can feed the next skill.
• Remaining questions are implementation details for the planner or executor.
• The user says "enough", "write it", "make the PRD", or equivalent.
• A blocker requires a separate research pass, stakeholder decision, or external access.

Final Interview Brief

End with this structure:

## Interview Brief: <topic>

### Context Read
- <files, issues, docs, or links used>

### Problem And User
<who has the problem, where it appears, and why it matters>

### Desired Outcome
<what must be true after the work ships>

### Version-One Scope
- <included behavior or decision>

### Non-Goals
- <explicitly excluded behavior or decision>

### Constraints And Dependencies
- <technical, business, timing, data, security, or UX constraints>

### Acceptance Signals
- <reviewable or testable completion signal>

### Risks And Open Questions
- <unresolved item, or "None">

### Recommended Next Step
<prd-writer | feature-intake | shape | spec-first | direct implementation>

Keep the brief concise enough to paste into a tracker issue or hand to a PRD writer. Include inference notes when a fact came from repo context rather than direct user confirmation.

Tell the user to run the recommended next skill. Do not fire it.

Anti-Patterns

• Dump a long questionnaire before reading repo context.
• Turn the interview into a PRD unless the user asks.
• Ask questions whose answers are already in .agents/memory/, root agent files, docs, code, or tracker context.
• Save plans in local agent plan files.
• Start implementation during the interview.
• Invoke another user-invoked skill from this one.