Rephraser & Feedback Loops

Expert

Use a secondary prompt to rewrite and clarify user requests before execution.

01 The Concept

Users often write vague prompts. Rephraser loops use a lightweight model to rephrase the user's intent into a structured, descriptive prompt, check if the user agrees, and then route the refined instructions to the main model, significantly improving success rates.

02 Weak vs. Strong

EX 01Rephraser Pipeline Setup

1. Expand query

expanded = call_llm(model="gpt-3.5-turbo", prompt="Expand query 'make logo' into a detailed graphic brief listing style, subject, and color scheme.")

2. Confirm description

user_approved = ask_user("Is this what you want? " + expanded)

3. Execute generator

if user_approved: call_llm(model="midjourney-agent", prompt=expanded)

→ Why it works

Vague instructions yield generic outputs. Expanding the brief first ensures the specialist model receives high-quality parameters.

03 Key Points

01Query Expansion: Turning short queries into descriptive task descriptions.

02Intent Refinement: Identifying implicit parameters (target audience, length) in short prompts.

03Verification Loop: Showing the expanded query to the user for confirmation.

04Specialist Routing: Selecting the best templates based on the expanded intent.

05Bypassing Vague Inputs: Preventing garbage-in-garbage-out failures in pipelines.

04 Model-Specific Notes

Claude is highly descriptive and makes an excellent query expander, generating detailed templates for downstream tasks.

05 For Your Role

It's like a translator who takes your short, casual requests and writes a polite, formal letter on your behalf.

Next: Context-Aware Summarization Loops →