Stage 02 of 11 · Prompt Engineering · 4 min · Reviewed Aug 2026
Prompt Engineering: System Prompts, Examples, and Control
A pasted persona is a starting point. Long-term control comes from a system prompt with a job, constraints, and output shape; from few-shot examples that show the boundary; and from formats the decoder can be held to. The library on this site collects templates. The workshop maps compilers, evals, and grammars. This lesson is the craft both of them assume.
- System prompts
- Examples
- Seed crystals
- Formatting
- Templates vs engineering
What you learn
How to reliably control model behavior without pretending a gist is a program.
The library — templates
A template is a static instruction: act as a role, fill slots, obey a reply shape. The public-domain recipes and system-prompt collection on this site live here. They are educational and pasteable. They have no metric, no trainset, and no decoder grammar unless the text happens to mention one.
Use a template for a one-off chat, a documented voice, or a demo. Do not use a template as the only control when the wording must survive model updates or when outputs must be valid JSON.
System prompts are job specs
The system prompt is the stable contract: who the model is for this run, what it may do, what it must refuse, and how it should format work. Put invariants here — not the user’s latest file. Keep it short enough to cache and audit.
A good system prompt names tools, stop conditions, and honesty rules (“if the tool failed, say so”). A bad one is a novel of vibe. If you cannot diff the system prompt in git, you do not own the agent.
You are the refund clerk for Acme.
Tools: get_order, request_approval.
Stop when the schema is valid or approval is pending.
If a tool failed, say so. Do not invent an order.Visual
System Prompt as a Contract
The system prompt defines the model's role, allowed tools, stop conditions, and output format. Together with tool schemas, it forms a stable contract that can be versioned, diffed, and audited.
Examples beat adjectives
Few-shot examples teach the boundary better than “be concise and professional.” Show one good output, one refusal, and one recovery from a tool error. Keep examples short; they compete with the real task for context.