---
title: "Grow the set without user text"
lesson_id: "05"
---

# Grow the set without user text

- **Lesson ID:** 05
- **Goal:** Grow the dataset with SampleGenerator and reviewed fixtures. Keep raw user text out of the default eval set.
- **Human lesson:** [05-grow-the-set.html](05-grow-the-set.html)

## Prerequisites

- Lesson 04, or a CI gate.

## Inputs, outputs, and artifacts

- **Inputs:** A small golden set.
- **Outputs:** A larger reviewed set and a privacy rule.
- **Artifacts:** Progressive `.flow` diagrams.

## Agent build steps

1. Name SampleGenerator and reviewed rows.
2. Forbid raw user transcripts in the default repo set.
3. Teach the flywheel: real use to production data to reviewed evals to a better agent.
4. Show a thumbs-down to synthetic fixture path. Traces are receipts. Link the tracer course.
5. Prefer harness, prompt, or tool fixes first. Fine-tune later.
6. Say the bottleneck is often writing down what good looks like.
7. Link the data flywheel course.
8. Pin prompt versions on samples.

## Constraints

Keep the language simple, friendly, and literal. iOS 27 and macOS 27 only. No quizzes, Mermaid, or em dashes. Do not invent API names. Mark sketches as sketches and tell the reader to confirm them in the Xcode 27 SDK.

## Key concepts

- `SampleGenerator` proposes more rows from seeds.
- Review every generated row. Keep the rejects.
- The default rule is no raw user transcripts in the repo or CI.
- A thumbs-down can become a fixture with a synthetic rewrite. Traces are receipts for the wrong step.
- Fix the harness first. Fine-tune later. The bottleneck is often writing down what good looks like.

## Takeaways

- Grade behavior, gate the mean, grow with review, keep raw user text out.
- Turn production misses into reviewed fixtures. Read traces to find the wrong step.
- Pin prompt versions on every sample.
- A rolled-back prompt should make the same eval go green.

## Acceptance checks

- SampleGenerator is named.
- Raw user text is excluded by default.
- Sibling links, progressive `.flow` diagrams, and an Agent brief link are present.