Claude Fable is the most powerful model out right now, and it is also the most expensive to run, roughly twice the cost of the next best model. That is fine when it builds the right thing. It hurts when it guesses wrong, hallucinates half a project, and you have to throw it all away, because you just paid premium tokens for something you cannot use.

There is one prompt that fixes this, and it takes about ten seconds to use. It is called the Interview Prompt. You paste it in before you start any project, and instead of diving straight in and guessing, Fable stops and interviews you first. It asks a set of questions that each attack the build from a completely different angle, then it writes a clear plan and waits for your go before it touches a single line of code.

The result is that Fable builds the exact thing you pictured on the first try, instead of burning your usage on the wrong one.

The Interview Prompt

This is the whole thing. Copy it and keep it somewhere you can grab before every build.

The Interview Prompt
Before you write any code, plan, or file, you must interview me. Do not start building until I have answered your questions and approved the final spec.

Ask me all of the following questions at once, numbered, in a single message. Each attacks the project from a different angle. Then wait for my answers.

1. OUTCOME: What are we building, in one or two sentences, and is it new or a change to something that already exists?
2. USER: Who is the primary human using this, and in what situation do they arrive?
3. THE ONE JOB: If this does exactly one thing right on day one, what is that thing?
4. DONE: What does "v1 shipped" look like, what's live, and where?
5. STACK: What technologies must I use, what am I free to choose, and what already exists (repos, databases, APIs, designs)?
6. CONSTRAINTS: What's the budget for third-party services, and what's the deadline?
7. FAILURE: What breaks the business if it fails silently, and what data can never be lost or leaked?
8. UGLY INPUT: What's the worst thing a user will do to this, bad data, weird behavior, abuse?
9. SCOPE CUTS: What might I assume is in scope that actually isn't? What's "later," not "now"?
10. HALF: If we could only ship half of this, which half?
11. TASTE: Describe or link something with the vibe you want. What do you hate seeing in things like this?
12. HARD RULES: Any non-negotiables about UI, naming, structure, or code style?

Rules:
- After I answer, if anything is vague or two answers contradict each other, ask targeted follow-ups before writing the spec. Do not assume.
- "I don't know" is a valid answer, so pick a sane default and flag it in the spec.
- Do not flatter my idea. If something is a bad plan, say so now, not after you've built it.

When you have what you need, write me a single spec containing: the goal in one sentence, the exact v1 feature list, the explicit non-goals, the stack, the file structure you plan to create, and the build order. End with: "Reply 'approved' to begin, or tell me what to change."

Do not write any code until I reply "approved."

How to use it (30 seconds)

You do not run this after you start. You run it first, before you have described anything.

  1. Open Claude and click Code in the top left, then start a new chat.
  2. Before you type anything else, paste the Interview Prompt into the box and hit send.
  3. Once it is done thinking, it comes back with all its questions in one message. Answer them honestly. If you do not know one, say so, and it picks a sane default and flags it.
  4. It writes you a full spec: the goal, the exact v1 features, what is out of scope, the stack, the file structure, and the build order. If it looks right, reply approved and it builds. If not, tell it what to change first.

Why twelve questions instead of just building

Most wasted tokens come from Fable quietly filling in the gaps you never spelled out, then building on top of a wrong guess. The prompt drags those gaps into the open before any code exists. Each question is aimed at a different way projects go sideways:

  • Outcome, The One Job, Done: pin down what you are actually building and what "finished" looks like, so it does not over-build or wander.
  • User and Taste: who it is for and the feel you want, so the result is not generic filler.
  • Stack and Constraints: what to use, what already exists, the budget, and the deadline.
  • Failure, Ugly Input, Hard Rules: what can never break, what a bad actor might throw at it, and your non-negotiables.
  • Scope Cuts and Half: what is "later" and not "now," so v1 stays small enough to actually ship.

There is one rule in there most people would never think to add: it is told not to flatter your idea. If your plan is bad, it says so now, while changing course is free, instead of after you have paid to build the wrong thing.

The trick is not a smarter model. It is making the model understand the job before it starts. Ten seconds of questions up front saves you a whole session of wrong turns.

What you actually save

Two things. First, tokens, because Fable is not writing, deleting, and rewriting whole files to chase a target that keeps moving. Second, your sessions, because you stop hitting your usage limit halfway through every build. On a model this expensive, that gap adds up fast, and it is the difference between shipping and staring at a wall until your limit resets.

It also works on more than code. Any time you are about to hand Fable something big and fuzzy, run the interview first and let it think before it acts.

Paste it in before your next build and watch how different it feels. One prompt, and Fable finally builds what you actually meant.
– Anir

If this is the kind of AI thinking you want working on your own business, apply to work with me. And to catch the next breakdown, join the newsletter.