Skip to content

How to Get AI Coding Assistants to Ask Clarifying Questions Before Generating Code

Problem

When I asked Codex to implement a feature, it would immediately dump a complete plan with all the wrong assumptions:

What Codex gave me
Here's my implementation plan:
1. Create a new UserService class with dependency injection
2. Add caching layer using Redis
3. Implement rate limiting
4. Create REST endpoints with JWT authentication
5. Add comprehensive logging
6. Write unit tests with 80% coverage

I never asked for caching. I never mentioned Redis. I didn’t need JWT auth. And now I’m stuck deleting code I don’t understand, trying to explain to Codex what I actually wanted.

The revision cycles would repeat endlessly. Each attempt, Codex would add more “helpful” features I never requested.

What Happened?

I found a Reddit post that captured exactly what I was experiencing. Someone posted about trying a skill called “grill-me” and getting a 90% success rate with Codex planning.

The core issue: AI coding assistants default to generating immediately rather than clarifying first. They fill gaps with guesses, add features you didn’t ask for, and create context mismatches.

The Solution

The grill-me skill is tiny - just 5 sentences:

grill-me skill content
Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.
Ask the questions one at a time.
If a question can be answered by exploring the codebase, explore the codebase instead.

This pattern changes the AI’s behavior from “generate immediately” to “interview first, then implement.”

How It Works

I tried the skill. Instead of dumping a plan, Codex now asks:

Codex interview session
Codex: Should the user service handle authentication, or should that be
a separate concern? I recommend keeping them separate - auth
changes frequently and you don't want user management coupled to
it.
Me: Separate makes sense.
Codex: Good. For user storage, should we use the existing PostgreSQL
database or create a new one? I see from your codebase that
you're already using Postgres for the product catalog.
Recommended: use the same instance with a new schema.
Me: Same instance is fine.
Codex: For the API layer, your codebase uses FastAPI. Should I follow
the same pattern, or did you have something else in mind?

The dialogue feels natural. Codex surfaces context I hadn’t considered. Decisions get made collaboratively. And when it finally generates code, it matches what I actually wanted.

Why This Works

The skill has three key components:

“Interview me relentlessly” - Sets the mode. The AI knows it’s in questioning mode, not generating mode.

“Ask the questions one at a time” - Prevents the AI from dumping 20 questions at once. This creates actual dialogue.

“If a question can be answered by exploring the codebase, explore instead” - Reduces redundant back-and-forth. The AI can self-answer many questions by reading your code.

Here’s what the behavior change looks like:

Behavior comparison
┌─────────────────────┐ ┌─────────────────────┐
│ Without grill-me │ │ With grill-me │
├─────────────────────┤ ├─────────────────────┤
│ Prompt → Generate │ │ Prompt → Interview │
│ ↓ │ │ ↓ │
│ Assumptions │ │ Clarify │
│ ↓ │ │ ↓ │
│ Revision Loop │ │ Generate │
│ ↓ │ │ ↓ │
│ Frustration │ │ Success │
└─────────────────────┘ └─────────────────────┘

Alternative Approaches

If you don’t have the grill-me skill installed, you can achieve similar results with inline prompts:

Alternative prompts
"Brainstorm with me about this feature before implementing"
"Let's discuss the design first. Ask me clarifying questions one at a time."
"Before coding, interview me about the requirements. Make recommendations for each question."

One Reddit user mentioned a ChatGPT pre-refinement workflow:

  • Discuss and refine the idea in ChatGPT
  • Ask ChatGPT to generate a prompt for Codex that includes all decisions
  • Paste that prompt into Codex

Creating Your Own Skill

I created a simplified version for my workflow:

SKILL.md
---
name: clarify-first
description: Ask clarifying questions before generating code. Use when planning features or making architectural changes.
---
# Clarify First Skill
## Instructions
Before implementing anything, interview me about my requirements.
1. Ask questions one at a time
2. For each question, provide your recommended answer with reasoning
3. Explore the codebase to answer questions when possible
4. Continue until we reach shared understanding
5. Only then, generate the implementation plan
## When to Use
- Adding new features
- Refactoring existing code
- Architectural decisions
- Bug fixes where root cause is unclear

The Result

After using grill-me for a few weeks, I noticed:

  • Fewer wrong assumptions
  • Less unwanted work to delete
  • Easier code review
  • Actually saving tokens through fewer revision cycles

Someone in the Reddit thread asked about token usage: “Doesn’t all that back-and-forth cost more tokens?” The answer was counterintuitive - it saves tokens because you avoid revision cycles. One revision cycle costs more than a focused interview.

Summary

In this post, I showed how the grill-me skill pattern changes AI coding assistants from immediate generators to thoughtful interviewers. The key insight is that a tiny 5-sentence prompt can fundamentally change AI behavior. For complex features where assumptions matter, this approach prevents the revision-cycle hell that many of us experience.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!

Comments