Prompt workshop
What is Textual Gradient?
A textual gradient is natural-language advice about how to change a prompt or other text parameter to reduce a language loss.
TextGrad uses this as an update signal. It is an inspectable critique, not a window into private chain-of-thought.
Why it matters
A textual gradient is natural-language feedback about how to change a prompt to improve a metric. Instead of numeric gradients (like in neural network training), TextGrad uses the model's own critique to guide prompt optimization.
This makes prompt optimization interpretable — you can read the gradient and understand what the optimizer wants to change.
Key takeaways
- 1Textual gradients are human-readable optimization signals.
- 2They make prompt optimization interpretable.
- 3The model critiques its own output to suggest improvements.
Common mistakes
- ✕Applying textual gradients without a clear metric to optimize.
- ✕Over-optimizing on training examples at the expense of generalization.
Related terms
Concept neighborhood
Terms linked from Textual Gradient in the glossary graph.
- Textual Gradient
- TextGrad
- Prompt Engineering