How I Built a Self-Correcting Prompt Generator with Multi-Stage LLM Calls"

💬 I don’t really know Python.

My background is Pascal, VB, and Prolog — structured and logical, but far from modern languages.

This system was built under pressure, pushed forward by Cursor and Copilot.

This post is about how I designed a self-correcting image prompt generator using a multi-stage LLM flow.

🎯 Why One Prompt Isn’t Enough

Most LLM prompt systems work like this:

1. You give it a caption


2. It generates tags or scene descriptions


3. You hope the structure makes sense

But most of the time, it doesn’t.

In a system like RΞNE — where consistency matters (we feed prompts to Stable Diffusion) — I didn’t want to manually audit outputs. So I built /genimgprompt: a self-checking, retry-capable multi-stage pipeline.

🧠 Step 1: Caption → Scene (Creative Persona)

We begin with caption + emotion + character + style_hint. The first stage uses a creative persona (RΞNE) to generate a vivid scene.

Example input:

{

  "caption": "A cyberpunk girl walks through a neon alley",

  "emotion": { "primary": "MYSTERIOUS", "intensity": 0.7 },

  "character": { "desc": "A mysterious girl with cybernetic eyes" },

  "style_hint": "cyberpunk, neon, atmospheric"

}

Scene result:

“She walks alone through the flickering neon-lit alley, her cybernetic eyes faintly glowing, casting light onto the damp ground. The atmosphere is saturated with mystery, her silhouette melting into the blurred textures of the cyberpunk city.”

🧩 Step 2: Scene → Structured Tags (Tagify)

Next, we ask the LLM to extract 6 structured fields:

• character


• pose_action


• outfit


• emotion


• background


• camera

Then we parse:

If the response fails parsing, we trigger retry logic:

• Retry with AI_Assistant persona (more structured)


• Still invalid? Use a default fallback

🔁 Retry and Fallback Mechanism

Here's the logic:

try:

    response = call_llm(persona="RΞNE")

    tags = parse_tagify(response)

except ParseError:

    response = call_llm(persona="AI_Assistant")

    tags = parse_tagify(response)

    if not valid(tags):

        tags = default_prompt_tags()

Why this matters:

• LLMs are unreliable no matter how strict your prompt is


• Any slight formatting deviation breaks downstream logic


• Debug info is logged at every stage

🧰 What I’m Open Sourcing

I’ll open source a simplified FastAPI module with:

• Input: caption, emotion, character, style_hint
  


• Output: prompt_tags, positive_prompt, negative_prompt
  


• Includes full debug logs of retries, LLM calls, and fallback triggers


• No memory system, no Stable Diffusion dependency, no image output

This module is fully standalone and can be integrated into any AI generation pipeline.

💭 Final Thoughts

This module isn’t about perfect code — it’s about a system that fails gracefully, recovers, and tells you what happened.

Side note: using AI/LLMs for coding is one of the best things that’s happened to me.

📡 Follow @n40-rene.bsky.social

Next post: full code and open-source repo release.