Skip to content

Module 01 / Lesson 01

Stop Vague Prompting with a Reverse Interview

Module 01 · Real-World ContextStep 01 of 03

Never ask an AI to write code until you have defined what success and failure look like.

Imagine your payment or messaging service degrades during peak hours. You need your team to publish an urgent incident notice to customers. If you ask an AI coding agent with a vague prompt like "build a status page with a publish button", it quickly outputs a flashy mockup with billing tabs and animations you never requested, but clicking Publish does not actually save or update anything.

Narrated lesson · 1:45
Read the narration transcript

A good-looking screen can hide a missing rule

Friday afternoon. I ask the agent for a service status dashboard. It gives the status app a timeline, a billing tab, and a Publish button. The preview looks finished.

Now I publish a note and refresh. The update is gone. I never said who could publish, who could read it, or what a successful save meant. The agent guessed.

Run the reverse interview

On Antigravity, I can install Superpowers using its repository URL. In a plain chat I can run the same reverse interview myself. The point is to stop before picking a stack or touching code.

I start with two questions. Who can publish? The operator for this organization. What can the visitor read? Approved updates, never the private troubleshooting note.

Then I ask where it saves, what happens if saving fails, and how we prove it worked. We write the answers into Spec dot M D. Billing and a separate realtime server stay out of this change.

Turn the answers into one small change

The issue is to publish one update on an existing incident. Read the existing membership rules and data types. Use the current public page. Ask if a storage decision is missing.

Keep billing and live transport out of this change. A realtime connection might make sense later. For this first slice, a refresh showing the saved update is already useful evidence.

Decide what the reviewer should see

Review as the operator first. Publish a sample update and refresh. The same text should still be there.

Then open the public page in a separate visitor session. Check that the approved text appears and the private notes stay out.

Finally, simulate a failed save in the local fixture. The draft should still be in the form. These are the checks that turn a vague request into a reviewable change.

The everyday problem

It’s Friday. You ask a coding agent to build a service status dashboard. The preview has a billing tab and a Publish button, so it looks close. You type an update, refresh the public page, and the old message is still there. Before asking the agent to fix it, ask what you meant by Publish. Should it save immediately? Who is allowed to do it? Who can read the result?

This lesson follows one incident on the Event Delivery service. By the end, you should be able to turn that vague dashboard request into a one-page brief another developer can implement and a reviewer can verify.

Why the agent got it wrong

The prompt named a product and left its behavior open. The agent filled the gaps with familiar dashboard features. Billing and a realtime feed are plausible guesses, even though neither solves our immediate problem. A polished preview can conceal those guesses because it can render sample data and report success without any persistent write.

A coding agent can produce a useful first draft once you give it the decisions it needs. Product intent supplies the criteria: an actor, an action, an observable result, and a result when the operation fails.

The tool and technique

Use the brainstorming skill in Superpowers to run a reverse interview. You give the goal and let the agent ask questions before it writes code. Start with five decisions: who publishes, who reads the update, where it’s stored, what a failed save does, and what evidence ends the task.

Answer one question at a time. If the agent asks whether we need subscriber billing, the answer can be no. A non-goal is an explicit boundary on the change. Writing it down makes it easier to reject extra dependencies and features when you review the patch.

Start with a reverse interview

Use the brainstorming skill in Superpowers to let the agent ask questions before it touches code. Decide who can publish, what goes on the public page, and what a failed save should do. Then use its writing-plans skill to turn those answers into a small implementation plan. GitHub Spec Kit is another option when the team already uses structured specs.

# Antigravity CLI, after reviewing the source and requested capabilities
agy plugin install https://github.com/obra/superpowers

Ask me one question at a time about operator access, visibility,
drafts, failed saves, and the proof of success. Ask before assuming.
Write the answers and non-goals in SPEC.md. Stop before editing code.

SPEC.md excerpt
Goal: An operator publishes one update on an existing incident.
Check: A public visitor sees the same text after refreshing.
Existing records: organizations, organization_members, services,
incidents, and incident_updates. Review their actual fields first.
Non-goals: Billing, subscriber alerts, and a new realtime server.

If the spec says “published” without naming what the visitor should see, the brief still needs a decision. For a typo fix, a short instruction can be enough.

On screen in the status app

The first request in the example leaves all five decisions to the agent. In the revised SPEC.md, a member with the operator role publishes on an existing incident. A visitor can read approved status updates. The editor retains the draft if the write fails.

Read the current schema before agreeing to add one. In the sample, organizations own services, a service has incidents, and incident_updates records their messages. organization_members tells us who can act for an organization. These relationships give the agent a route to inspect without guessing a new database or inventing a realtime server.

SPEC.md

Before

Build a real-time service status dashboard with a publish button. Make it production-ready.

After

Goal: an authorized operator publishes one update on an existing incident.
Data: use organizations, organization_members, services, incidents, incident_updates.
Public: show approved incident messages. Keep internal notes private.
Failure: retain the draft and show a retryable error when the write fails.
Verification: a separate visitor sees “Recovery in progress” once after refresh.
Non-goals: billing, subscriptions, and a new realtime server.

What to watch for

Superpowers can structure the question-and-answer process. It cannot decide the right behavior for the business or confirm the repository matches the spec. Read the spec against the files and ask the owner to resolve any unresolved permission decision before implementation.

Match the effort to the uncertainty. A typo in a field label can fit in one sentence. Editing who can publish or what becomes public deserves a written decision and an explicit stop before the agent starts changing files.

How you know it worked

Read the acceptance check like a script. Publish “Recovery in progress” as a seeded operator, open the public page separately, and reload it. The message should persist and appear once. With a seeded failing write, the editor should keep the draft and display an error.

Ask a reviewer to point to the sentence that authorizes each UI or schema change. If no sentence supports billing, skip that diff. The next lesson shows how to keep the database-client and permission conventions available to a fresh coding session.

References