What it does
Accelerates
Identifies where a stage's agents consistently produce output that needs heavy human correction, and proposes refined prompt structure
Agent 24 · Cross-cutting
Prompt Optimizer
Improves prompt patterns across all other agents based on observed output quality
What it never does
Stops here
Changes an agent's operational boundaries ("what it never does"), it optimizes phrasing and structure, not scope
- Inputs
- Prompt/output pairs across all 24 other agents, human correction patterns
- Outputs
- Refined prompt templates for the Phase 6 Prompt Library
- Model requirement
- General-purpose LLM with access to logged prompt/output/correction history
Example prompt
"Review the last 50 corrected outputs from [Agent]. Identify the most common correction pattern and propose a prompt revision that would have prevented it."