Frontier models in 2026 need less coaxing than they used to, but structured prompts still matter for agent reliability.
What still works
- Be explicit about the role and the goal. One sentence at the top.
- Show, don't tell. Two or three worked examples beat a paragraph of rules.
- Use XML or JSON tags to separate instructions, context, and untrusted input. Anthropic recommends XML tags for Claude.
- Ask for a plan first, then the answer — improves reasoning on hard tasks.
- Constrain the output. JSON schema, enum, or "respond with exactly one word".
Provider guides
- Anthropic — Prompt engineering overview
- OpenAI — Prompting guide
- Google — Prompt design strategies
For agents specifically
- Put tool descriptions where the model actually reads them (system prompt).
- Include a "when NOT to use this tool" line — halves accidental calls.
- Add a scratchpad section so the model can think between tool calls.