Interview Mode: Let Claude Code Ask the Questions First
Most developers start wrong. You fire up Claude Code, paste a prompt, and hit enter. Claude makes assumptions. Lots of them. By the time the code appears, you realize you wanted OAuth instead of sessions, or a third-party auth service instead of rolling your own.
Then you rework everything.
Spec-driven development flips this. Let Claude ask the questions first.
The Problem With One-Shot Prompts
When you ask Claude to "add authentication to my app," it has to guess. Is it a SPA? Mobile app? What's your auth strategy? JWT? Sessions? OAuth? Do you need multi-tenancy? Should you use a managed service like Clerk or WorkOS?
You didn't specify. Claude didn't ask. It shipped code based on assumptions that were cheap to change before being built, but expensive after.
This is the hidden cost of prompt-driven development: you're making critical architectural decisions implicitly, discovering them later during code review when fixing them means throwing away tokens and time.
Interview First, Spec Second, Code Last
The antidote: let Claude interview you.
This idea, shared by Tariq at Anthropic, is straightforward: instead of guessing what you want, Claude uses the Ask User Question tool to drill into requirements. Not obvious questions—deep ones.
One developer reported Claude asked 40+ questions before finalizing the spec. 40 questions they never would have answered upfront, but that made the spec bulletproof.
The workflow looks like this:
- Provide a minimal prompt ("I'm building a Next.js marketing site for developers")
- Let Claude interview you using Ask User Question tool—technical decisions, UI/UX concerns, trade-offs
- Output a detailed spec (not code)
- Start a new session to execute against that spec
This forces decisions to the surface when they're cheap to change.
How It Works in Practice
You create a skill that triggers the interview automatically. The prompt is simple:
"Read the spec.md and interview me using the Ask User Question tool about technical implementation, UI and UX concerns, trade-offs. Make sure questions aren't obvious. Be in-depth and continue until complete. Then write the spec to the file."
Claude asks. You answer. It synthesizes into a formal spec. No code yet.
This is not a replacement for Plan Mode (which you should still use). Think of interview mode as the precursor to planning—nail requirements first, then plan implementation.
Why This Actually Saves Time
Counterintuitive: slowing down speeds you up.
The longer you spend planning, the less time reworking. Because you're narrowing the solution space before Claude burns tokens generating code.
Instead of discovering buried assumptions during code review, you confront them when they're cheap to change. Instead of you guessing and Claude correcting, Claude asks clarifying questions instead of making assumptions.
This is a fundamental shift in how agentic AI works. Traditional prompt engineering demanded perfect instructions upfront. Spec-driven development lets AI help you discover what you actually want—because you probably don't know all the nuances before talking it through.
The Real Win
You get control back.
Most AI coding tools work top-down: you specify, they build. Here, it's bidirectional. Claude doesn't assume. It asks. You don't have to guess. You decide.
For large features, this changes everything. For a complex auth system, CMS integration, or multi-tenant setup, the difference between building once and building twice is hours of wasted effort.
Next time you have a large feature, try it. Don't cram everything into one prompt. Let Claude interview you. You'll be shocked how many requirements you didn't even know you had.
Further Reading
- Tariq's original insight: Search "spec-driven development Claude Code" on Twitter for the original thread
- In Claude Code: Try creating a skill that triggers interview mode—the Ask User Question tool is built in and designed for exactly this workflow
