Exploring a more deterministic approach to AI-assisted code generation

[MyrinNew]

Introduction

AI coding agents are getting surprisingly good.


In small projects, you can ask them to add features, fix bugs, and even write tests—and they often succeed.


But once your project grows, things start to break down.


In my experience, the issue is not model capability. It’s something more subtle: prompt instability.


The Problem: Prompt Instability

Most coding agents construct prompts dynamically using:

• chat history
• parts of the codebase
• internal heuristics

This means the final prompt is not fully under your control.


As a result:

• the same request can produce different outputs
• changes can appear in unexpected parts of the codebase
• behavior becomes harder to reason about

In small projects, this is manageable.

In larger systems, it becomes risky.

A Different Approach: Treat Prompts Like Source Code

Instead of relying on dynamically constructed prompts, I started experimenting with a different idea:

Treat prompts like source code.

That means:

• prompts are explicit
• prompts are reusable
• prompts can be composed from other prompts
• prompt construction is deterministic
• prompts are the source of truth, not code

This shifts the workflow from “chatting with an agent” to something closer to designing system architecture.

The Tool: SVI (Structured Vibe Coding)

To explore this idea, I built a small tool called SVI.


SVI generates source code from structured specification files (.svi) written in a Markdown-like format.


Key ideas:

• Each .svi file defines how a specific source file should be generated
• Prompts can import and reuse other prompts
• The final prompt is constructed in a fully controlled and predictable way

Unlike typical coding agents, SVI does not depend on chat history or implicit context.

Example

Here is a simple .svi file:






# Destination File
hello.js

# Output
function hello()

# Options
ProgrammingLanguage=Node.js
Active=True

# Prompt
Create a function that prints "Hello World", and call this function








Generate the code with:






svi run







The output is generated by an LLM and based on the specification.


Why This Matters

This approach has a few practical benefits:

• More predictable results

You know exactly which prompt generated which file, and get more predictable results

• Reusability

Prompts can be shared and composed

• Lower model requirements

Smaller prompts allow you to use cheaper or even free models; you can adjust the prompt size and complexity to match the LLM you’re using.


Trade-offs

This approach is not a silver bullet.

• It requires more upfront structure
• It is less flexible than free-form prompting
• It changes the workflow from interactive to more declarative

However, for larger projects, this trade-off might be worthwhile.


Conclusion

AI coding agents are powerful, but their current design makes them hard to control at scale.

Treating prompts like source code is one way to bring back structure and predictability.

I’m still experimenting with this approach, and I’d be interested to hear if others have explored similar ideas.


Links

GitHub repository: https://github.com/avrmsoft/svi




More...