Skip to content

Learning path · Programmatic Prompting · 31

Teleprompters

DSPy optimizers that search over prompts, few-shot sets, or module parameters to maximize a training metric.

Why it matters

  • Automates prompt iteration with reproducible experiments.
  • MiPROv2 and GEPA represent different search strategies.
  • Requires trustworthy eval signals to avoid overfitting.

Key ideas

  • Search space
  • Metric-driven selection
  • Bootstrap demonstrations

Teleprompters treat prompt design as search. They mutate instructions, select demonstrations, or adjust combinational rules, keeping changes that improve dev-set scores. Guard against overfitting tiny sets—hold out production-like failures. Combine with human review of winning prompts; optimizers exploit judge blind spots. Version optimized artifacts like any model checkpoint. Hold out a production-like failure set untouched during search; promote winners only when they generalize beyond training quirks. Ship only after eval gates pass on representative production failures.

Updated 2026-08-09 · Full learning path