OpenAI (GPT-4.1 / GPT-5 guides): start with Markdown titles per section (XML also performs well); in long context, place instructions at both the beginning and the end; separate data from instructions with delimiters; define identity and goals; for JSON, prefer Structured Outputs with a schema. Whether to ask for explicit step-by-step reasoning depends on the model — recommended for GPT-4.1, unnecessary for o-series reasoning models, and tied to reasoning effort on GPT-5.
Anthropic (Claude prompting best practices): structure with XML tags; put long documents at the top and the query at the end; give Claude a role in the system prompt; use 3–5 relevant examples wrapped in <example>; tell Claude what to do instead of what not to do; add the motivation behind instructions. On current models thinking is a model capability (adaptive, always on in the newest ones) — manual chain-of-thought is only a fallback.
Google (Gemini prompt design strategies): always include few-shot examples; use prefixes such as Input:/Output: or clear delimiters; specify constraints and response format; break complex tasks down; with lots of content, supply the context first and put the question at the very end.
This checker encodes only rules that are documented and verifiable in the text. Rules that depend on model configuration (reasoning effort, thinking) are shown as notes, not scored. Rules re-verified against the guides on 2026-08-23.
Frequently asked questions
What does this prompt format checker actually check?
It checks your prompt text against conventions documented in the official prompting guides of OpenAI, Anthropic and Google: structure (Markdown sections vs XML tags), where instructions sit in long context, explicit output format, delimiters between data and instructions, role/persona, examples, and positive vs negative instructions. Each rule links to the exact guide it comes from. It does not measure output quality — for that, use the full evaluation.
Is this real prompt engineering or just style preference?
It is conformance to documented guidance, not a performance guarantee. The providers publish these conventions because they measured them on their own models (for example, Anthropic reports up to 30% better responses when long documents go at the top and the query at the end). Effect sizes are modest and models keep converging, so treat the checker as a fast, free hygiene pass — and note where providers disagree (chain-of-thought is recommended for GPT-4.1, unnecessary for OpenAI reasoning models, and model-native on current Claude).
Does Claude really prefer XML tags?
Yes — Anthropic's guide says XML tags "help Claude parse complex prompts unambiguously" and recommends wrapping each content type in its own tag (<instructions>, <context>, <example>, <document>). OpenAI's GPT-4.1 guide recommends starting with Markdown section titles and notes XML also performs well. Gemini's guide recommends clear delimiters (XML or Markdown) and prefixes such as Input:/Output:.
Where should instructions go in a long prompt?
OpenAI's GPT-4.1 guide: put instructions at both the beginning and the end of long context. Anthropic: put long documents at the top and the query at the end. Google: supply all the context first and place the question at the very end. The checker applies these placement rules only when the prompt is long enough for them to matter (about 3k tokens and up).
Is it free? Do you store my prompt?
Free, no signup, no limit. The check runs entirely in your browser — the prompt is never sent to a server. The full evaluator (0–100 score, issues located in your text, rewrite) is a separate tool with a free tier.
Can it rewrite my prompt for a specific provider?
The free checker only reports. Inside PromptEval, paid plans include "Adapt for provider", which restructures your prompt to meet the failing rules for OpenAI, Anthropic or Gemini without changing what the prompt asks — and re-runs this same check before and after so you can verify the result.