Spec / 016 · workshop
A pasted persona is a starting point. Long-term value is treating the prompt as a program: compile it, test it, constrain generation, or search for better wording against a metric. Each method has a steppable conceptual workflow. This site does not run DSPy, promptfoo, Guidance, TextGrad, or GEPA live.
01 · Compile
Search instructions and demonstrations against a metric.
1 method
02 · Test / regress
Regression-test prompts so wording cannot silently drift.
1 method
03 · Constrain
Enforce structure in the decoder, not in a later apology.
1 method
04 · Optimize
Search or evolve wording automatically against a loss.
2 methods
How to read a method page
01 · Play
Step the workflow. Each stage is an observable artifact, not a hidden thought.
02 · Evidence
The spec panel highlights the sentence the active stage is drawn from.
03 · Compare
Method vs template shows what a library paste cannot do. Templates stay at /prompts.
Compile
Search instructions and demonstrations against a metric.
Test / regress
Regression-test prompts so wording cannot silently drift.
Constrain
Enforce structure in the decoder, not in a later apology.
Optimize
Search or evolve wording automatically against a loss.
TextGrad optimizes prompts with textual gradients: forward generation, a language loss, backward advice, update.
Hand-rewriting a library persona has no backward pass. You cannot point to the error and get a structured update.
Open method →
GEPA evolves prompts by reflecting on traces and keeping a Pareto archive — a population, not a single rewrite.
A single library persona cannot represent trade-offs. The short prompt and the accurate prompt are often different artifacts. Pareto search keeps both until you choose.
Open method →