# Test before you run

> Check a prompt on a sample before you enrich everything, and what to look for in the results.

Enriching a full dataset takes time. A few minutes checking the prompt on a sample first catches most problems.

## Two ways to test

- **Ask the agent to label a sample.** With **Prompt Designer** turned on in the agent's settings, ask: *"Pick 40 varied tickets and show me how this prompt would label them."* This is the agent applying your prompt in the chat, a preview, not a run of the enrichment itself. It's the fastest way to iterate.
- **Backfill a short range.** Save the prompt, then on the **Backfill** tab run it over one source and a few days. Those labels come from the real enrichment, so read them before you backfill the full history.

Do the first until the sample looks clean, then the second before a large backfill.

## Build a sample that covers the edges

Don't take the first 50 rows. We suggest 30 to 50 records across the file: the common kinds and the rare ones that matter (a cancellation, a complaint about safety, a one-word reply, a very long thread, a heavily redacted one).

## What to check

| Check | What goes wrong | Fix |
|---|---|---|
| Values outside your list | The model invents `neutral` or `general_mention` | Add "Do not use X; use Y" to the description |
| Quotes that aren't quotes | A verbatim field paraphrases instead of copying | Say "copy the exact words; do not paraphrase" |
| Label without a quote | A judgement is set but its verbatim is `none_detected` | Make the two descriptions refer to each other |
| Speaker-label leaks | A persona or brand Dimension fires because the transcript says "Maya (AI):" | Say "ignore speaker labels" |
| Contradictions | `resolved: true` with an outcome of `escalated` | Define the two together |
| Everything `none_detected` | The Dimension asks about something the text rarely contains | Check it is a text signal at all, then add phrases from your data |
| One value dominates | Nearly everything lands in one category | The categories are too coarse, or the description is too loose |

## Run, read, fix, repeat

Fix the descriptions, then test again on records you haven't used yet. Two or three rounds is normal. When the sample is clean, save the prompt and [backfill](/dimensions/backfill) it, then spot-check the real results the same way.