Evals before vibes: fixing a prompt that drifted

Problem

I own the production LLM system that generates curriculum-aligned lesson plans for teachers in Peru, Colombia and Mexico, running on Gemini 2.5 Pro through Vertex AI. A lesson plan is only useful if it follows the structure its national curriculum defines, section by section.

Measured against that structure in production, only 45.6% of the responses conformed. Some responses were also bloated, spending tokens on text nobody had asked for. The cure for that is not a better-sounding prompt. It is a way to measure one.

Approach

I built a set of test fixtures and an LLM-as-judge rubric that scores every output on 5 dimensions and 19 items. Each prompt change now produces a number that can be compared with the last one, instead of an impression.

1. 01Versiona number per change
2. 02Fixturesthe same set each run
3. 03Judge19 items in 5 dimensions
4. 04Release45.6% prod → 8/8 fixtures

1. 01Every prompt change got a semantic version, so it shipped as a release with a number rather than an edit in place.
2. 02The same set of test fixtures ran against each new version.
3. 03An LLM-as-judge rubric scored every output on 5 dimensions and 19 items.
4. 04Each score was compared with the last version's: structural conformance went from 45.6% in production to 8 of 8 test fixtures passing, and the dashboard showed what each version cost.

Step 1 of 4

Every prompt change got a semantic version, so it shipped as a release with a number rather than an edit in place.

Prompts became versioned artefacts with semantic versions, so a change is a release with a number rather than an edit in place. In the same system, the unit-generation engine I shipped keeps checks that do not need a model out of the model: deterministic guardrails validate the structure and detect truncated output without a single extra LLM call. And every call reports its tokens, cost and latency, so a change that improves quality but doubles the bill is visible on the dashboard before it becomes a surprise.

Result

Structural conformance went from 45.6% of production responses to 8 of 8 test fixtures passing after the fix.

The lasting change is in how prompts ship. A new version is a measured release rather than an opinion: the fixtures say whether it is better, and the dashboards say what it costs.