Use CALIBER
Prompts
The task-oriented entry point for prompt assets in CALIBER: author, evaluate, calibrate, release, and recover without starting from deep architecture.
This page is the task-oriented entry point for using prompt assets in CALIBER. Use it before the deep prompt architecture page when the question is practical: author, evaluate, calibrate, release, or recover.
At a glance
| Task | Start here | Deep reference |
|---|---|---|
| Edit a prompt safely | create or update a prompt revision | Prompts architecture |
| Measure prompt quality | run evaluation against a test set | Evaluation |
| Generate better candidates | launch prompt calibration | Calibration |
| Make a candidate live | use the explicit release/apply path | The refinement loop |
| Recover from release trouble | use the runbook | Operations runbook |
1. What prompts are for in CALIBER
Prompts are governed assets with version history, evaluation context, and an explicit release boundary. Editing a prompt does not make it live by itself.
That distinction matters because CALIBER separates:
- authoring
- evidence generation
- candidate generation
- live release
2. Common tasks
| You want to... | Read this next |
|---|---|
| create or revise a prompt | Prompts architecture |
| compare prompt behavior against known examples | Evaluation and test sets |
| generate improved candidates | Calibration |
| understand release, rollback, and review boundaries | Review and release flows |
3. What usually confuses readers
| Symptom | What it usually means |
|---|---|
| The prompt changed, but production behavior did not | the change is not live until the explicit release path completes |
| A candidate scored well, but nothing deployed | evaluation or calibration produced evidence, not live apply |
| Prompt behavior regressed after a release | inspect the release path, evidence set, and rollback options |