Prompt engineering is not a trick — it is a discipline. Once you internalize the underlying model, you stop guessing and start designing.
The 4 levers
- Role — who the model is pretending to be
- Context — what background it needs
- Task — the specific output you want
- Constraints — length, format, tone, what to avoid
Why "role" matters
Telling the model "You are a senior tax attorney" measurably shifts the vocabulary, caution level, and citation behavior. Roles activate clusters of training data.
Output formats
Always specify: Markdown table? JSON? Bullet list? Numbered with citations? Without this, the model defaults to wall-of-text — which is almost never what you want.
Reusable templates
Save 10 prompts that cover 80% of your work. Iterate on them weekly. This is exactly how power users 10x their throughput.