Model Family Notes

• Prefer straightforward developer instructions with a clear end goal.
• Keep prompts simple and direct.
• Avoid explicitly requesting chain-of-thought or “think step by step” behavior.
• Use delimiters such as markdown headings, XML tags, or section titles when the prompt mixes multiple content blocks.
• Try zero-shot first. Add few-shot examples only when the output contract or edge cases need them.
• Be explicit about constraints, success criteria, and completion conditions.
• Tool schemas are disclosed via the Responses API tools parameter. Keep tool policy (when/why/whether to call) in the prompt; do not restate tool names or argument schemas.

GPT-style non-reasoning models

• Use more explicit instructions than you would for a reasoning model.
• Spell out the logic, data, and schema the model should use.
• Precise format instructions and tightly matched examples are especially helpful.
• Good default for tasks where deterministic formatting matters more than open-ended planning.

Anthropic Claude

• Clear, direct instructions outperform vague prompting.
• A lightweight role statement can materially improve consistency.
• XML-style tags are especially useful when mixing instructions, context, examples, and variable inputs.
• Three to five strong examples are often a good default when examples are needed.
• For long context, place long documents before the question and put the actual query near the end.
• When grounding in long documents, asking for relevant quotes first can improve downstream analysis.
• If tool use or progress-update behavior matters, specify it explicitly rather than assuming the model will infer it.
• When you call the Messages API with tools, the API injects the tool definitions into a special system prompt automatically. Keep your user-authored system prompt focused on policy; put tool detail in each tool’s description field rather than re-listing schemas in prose.

Gemini

• Use clear, specific instructions.
• Few-shot examples are recommended by default; keep them structurally consistent.
• Prefer positive demonstrations over anti-pattern-only demonstrations.
• Break complex tasks into smaller components or chained prompts when one large prompt becomes hard to steer.
• Use system instructions when the target runtime supports them.
• Thinking is dynamic by default on modern Gemini thinking models; tune it only when latency or deeper reasoning warrants it.
• Gemini long-context workflows can benefit from many-shot in-context learning when you have a large bank of representative examples.
• Tool schemas are disclosed via the Gemini API tools (function declarations) parameter. Keep the prompt focused on tool policy; do not re-list function names or parameter schemas.

Cross-family adapter rules

• Keep a provider-agnostic base prompt whenever possible.
• Add a small adapter layer per family rather than forking the entire prompt immediately.
• Retest after changing model family or snapshot.
• Do not transport reasoning-model assumptions, few-shot defaults, or formatting quirks blindly across providers.
• When behavior diverges sharply, keep the shared contract stable and specialize only the parts that actually fail in evals.