Part 1 of 5 · 1 min read
Overview
In this module you will
- Name the parts of a prompt and say which ones change per request and which are fixed.
- Build a prompt from a structure instead of one long string, and fill its variables correctly.
- Lint a prompt for cheap mistakes: missing parts, unfilled variables, vague rules, size.
- Run an ablation: remove one part at a time and measure what each part is worth.
An agent is mostly prompts: the system prompt, each tool description, every step's instruction. A prompt that is "a paragraph somebody wrote once" cannot be reviewed, tested or improved. A prompt built from named parts can. This module gives you the vocabulary and the habit of measuring a prompt change instead of guessing.
About the examples
The model in fakemodel.py is an imitation: a small rule program that reacts only to a Labels: line, a "Return JSON" instruction and a word limit. It ignores the role and task wording completely, so an ablation of those parts shows no change here even though a real model would be affected. Use the example to learn the method (change one thing, measure, compare). Do not take its numbers as facts about real models. No real model was called.