Build high quality AI features with simple feedback loops (dovetail.com)

Dovetail’s AI team, led by Peter Wooden, argues that prompt testing should be as fast and visual as building UI components in Storybook. Facing more prompts than engineers, they rejected heavyweight hosted tools and end-to-end-only evals, instead treating prompts and agents like software components with clear contracts. The result: lightweight, repeatable “unit evals” and snapshot-style tests you can create in ~20 minutes, run in watch mode, and review inline in PRs. This approach closes the gap between shipping velocity and model quality—avoiding both rushed launches and analysis paralysis—and makes prompt changes easy to inspect, diff, and iterate on. Technically, the toolkit is simple: isolate a prompt, create a handful of code-based inputs, write scripts that emit outputs and diagnostics (including chain-of-thought), and visualize results as syntax-highlighted diffs. Start small (5→10→20→50→hundreds), use basic string matches and human-labeled ground truths, and add quantitative metrics (precision/recall) tracked in Git. Apply the testing pyramid—favor unit evals for prompt chains and use E2E only where necessary—spot-check LLM judges, and let user feedback guide which failure modes to measure next. The implication for AI/ML teams is clear: treat prompts as testable components, use fast visual feedback loops to find and fix failures quickly, and scale quality iteratively without sacrificing speed.