Prompting and Instructions¶
Two ways to steer a model: the prompt you send with each request, and the standing instructions that arrive with every request. This subtopic covers the craft of both — and the principles that decide what belongs where.
flowchart LR
J["Junior: the 5-part prompt"] --> M["Middle: examples, formats, iteration"]
M --> S["Senior: standing instruction files"]
S --> P["Professional: instructions as team asset"]
Levels¶
| Level | Guide | You are done when |
|---|---|---|
| Junior | The 5-part prompt | You can write a prompt with role, task, constraints, format, and examples — and know why vague prompts fail. |
| Middle | Examples, formats, iteration | You can use few-shot examples, structured output, and a real-input iteration loop to make behavior consistent. |
| Senior | Standing instruction files | You can apply the principles that make AGENTS.md/CLAUDE.md/SKILL.md files work — layering, brevity, verifiability. |
| Professional | Instructions as team asset | You can govern instruction files: ownership, review cadence, drift detection, and bloat control. |
Practice rule¶
Every instruction — in a prompt or a standing file — competes for the model's attention. If a rule isn't load-bearing enough to state precisely, it isn't load-bearing enough to include.
Related¶
- Context Engineering — the token budget every instruction spends from.
- Datasets and Graders — how to prove a prompt change helped instead of just feeling better.