Definition
Plain language
A method that improves an AI's prompts by treating written feedback like a correction signal.
As stated in the literature
A prompt-optimization framework that backpropagates natural-language critiques through a system of LLM calls as a textual analog of gradients; a baseline against optimize-anything and SkillOpt.
Why it matters: It improves prompts using plain-language feedback as a stand-in for training signals, automating a tuning step that's usually done by hand.
For example, it reads a written critique like 'your answer was too vague' and uses that feedback to rewrite and sharpen the prompt.
Heard on the show
“… And existing optimizers for these systems — TextGrad, DSPy, GPTSwarm, that whole lineage of treating an LLM pipeline as a program you can automatically …”Episode 181 — How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires