AI System Prompt Generator

A system prompt is standing guidance that shapes every reply an assistant gives. This tool does not call a model or invent wording — it arranges the text you type into the section order that Anthropic, OpenAI and Google all recommend, then wraps it in your chosen delimiters. The same inputs always produce the same output, so it is predictable and explainable rather than a black box.

The assembly is deterministic string composition:

1. Fixed section order. Role → Task → Context → Tone → Output format → Constraints → Examples. All three vendor guides state that instructions come before examples and that the output format should be stated explicitly, so examples are always placed last.
2. Role line. Your role becomes You are … — an indefinite article is added only when the role reads like a common-noun descriptor. If you give an audience, a You help … clause is appended. Assigning a persona is a documented best practice in the OpenAI and Gemini guides.
3. Tone and format become instructions.The tone you pick maps to a short directive (Concise → “Keep responses short and to the point; avoid filler.”), and the output format maps to an explicit instruction (JSON adds “Respond ONLY with valid JSON, no prose.”). Explicit format is called out in the OpenAI and Gemini guides.
4. Guardrails become bullets. Each non-empty line in the guardrails field is rendered as a bullet under a Constraints heading, so must-not-do rules are unambiguous.
5. Examples last.Each input/output pair is rendered after all instructions, matching Anthropic's and OpenAI's guidance to place reference examples at the end.
6. Style and length. The Claude style wraps sections in XML tags (<role>, <task>, …), as Anthropic recommends; Generic, OpenAI and Gemini use Markdown headings. Compact drops the tone block for brevity; Detailed adds a one-line rationale under each heading. Empty optional sections are never printed.
7. Token estimate. The character count is divided by four for a rough token figure (labelled approximate). A separate word-based cross-check runs internally, and the Token Counter gives an exact per-model count.

A built-in structure check independently confirms that every piece of text you entered — role, task, each guardrail, each example — actually appears in the output and in the correct order. When it passes, the card shows how many checks succeeded, so the result is verifiable rather than taken on trust.

The generator invents no rates or numbers. Its “source” is the documented prompt-engineering best practice from the three major model vendors, which fixes the section set and their order:

• Anthropic — Prompt engineering overview (role prompting, XML delimiters, examples last)
• OpenAI — Prompt engineering / GPT best practices (clear instructions, personas, delimiters, specify format)
• Google — Gemini API prompting strategies (assign a role, add constraints, specify output format, few-shot)

The section set and ordering were last cross-checked against these vendor guides on 2026-07-04. They are reviewed whenever a vendor materially updates its prompting guidance.