I tried this approach and found its not good for scale. Lots of areas where this approach will overflow max input tokens + model need to be conversational with coding capabilities atleast json.