Did the AI do what you asked?
You asked for under 120 words, no exclamation marks and only the facts you gave. Paste the prompt and the answer, and see every rule it kept or broke, the details it made up, and the lines that make it hold next time.
Read why AI ignores instructions, and how to word each rule so it holds. For JSON, see how to get valid JSON every time.
New to this? What it does, in plain words
It reads your prompt for the rules you set, then checks the answer against each one.
Rules like a word limit, a number of bullet points, a table with certain columns, JSON only, a subject line, no emoji or a sign-off name. AI answers break these quietly, and a quick read rarely catches it. It also lists the numbers, names and links in the answer that were never in your prompt, which is where made-up facts hide.
What you need
Nothing. No API key and no sign-up. When you press Check, your prompt and the answer go to thinkhard.si, are read, and are forgotten.
How to use it
- Paste the prompt you sent, then the answer you got back.
- Press Check the answer.
- Each rule shows Kept, Broken or Read it yourself. Point at one to see the words in the answer that decided it.
- Copy the prompt with the fixes added, or send the rules to the Prompt Tester so every new version is held to them.
Your prompt
It also had a system prompt or custom instructions
The AI's answer
ExampleWhy answers break the rules you set
A model writes one word at a time and keeps no running count, so a limit of 120 words is a target, not a ceiling. It is trained to be friendly, so it opens with a cheerful line and closes with an offer. And when a fact is missing, it fills the gap with something that sounds right.
- You. Paste bothThe prompt you sent and the answer you got.
- thinkhard.si engine. Find the rulesLengths, counts, formats, what to include and leave out, language, voice.Instant
- thinkhard.si engine. Judge each oneKept, broken, or for you to read, with the words that show it.
- thinkhard.si engine. Spot what was addedNumbers, names, links and quotes that were never in your prompt.
- You. Make it holdThe lines to add, and tests for the Prompt Tester.
What it checks
- Length. Words, characters, sentences, paragraphs and lines, for the whole answer or for each bullet, headline or subject line. Words are counted the way the Prompt Tester counts them, without the chatty first line and closing offer.
- How many. Three subject lines, five bullet points, ten ideas: it counts the list the answer actually gave.
- Format. A bulleted or numbered list, a table and its columns, JSON and its keys (and whether code could parse it as it came), comma-separated rows, named sections, plain text with no markdown.
- What to include and leave out. A subject line, a call to action, a sign-off name, words you named; and no exclamation marks, emoji, hashtags, long dashes, jargon, apologies or a phrase you banned.
- Voice and language. The language and the spelling you asked for, first person or "you", and whether the words are plain enough for the reader you named.
- Behavior. Only the answer with no introduction, asking questions before answering, and marking a missing fact instead of guessing.
What the answer added
Every number, date, name, link and quote in the answer is compared with your prompt and the material in it. Anything new is listed, because that is where invented facts hide: a statistic nobody gave it, a company name, a link that may not exist. Some additions are fine, such as a greeting with the reader's name. The list is there so nothing new slips past unread. To check claims against the world, use the Fact Check List.
It only judges what a rule can see
Whether an email is persuasive, or a summary fair, takes a reader. Those rules are marked Read it yourself, with what it could measure, such as the reading level or the first sentence. Instructions inside material you pasted, such as an email that says "reply by Friday", are not counted as rules for the AI, and neither are rules inside an example or ones that start with "if".
Questions
Is my text stored?
No. When you press Check the answer, the prompt and the answer go to thinkhard.si, are read by the engine, and are forgotten. Nothing is written to disk or logs, and nothing is used to train anything. The only record is a count of how many checks were run each day. See the privacy page.
Which AI tools does it work with?
Any. Paste the prompt and the answer from ChatGPT, Claude, Gemini, Copilot, Perplexity or a model you call from code. The fixes can be laid out for Claude, ChatGPT or Gemini, which each read prompts a little differently.
Why does it say my answer broke a word limit by only a few words?
Because it did. The count skips an opening line such as "Sure, here it is" and a closing offer such as "Let me know if you want changes", since you would delete those anyway. If a few words over does not matter to you, the verdict still tells you the limit is not holding, which matters once the prompt runs a thousand times.
Can it check more than one answer?
Yes, one at a time here. To hold a prompt to its rules on many inputs and models at once, press Test it in the Prompt Tester: the rules become checks there, and every run is scored against them.
Can I use it from my own code or from an AI assistant?
Yes. POST /api/v1/check-answer with the prompt and the answer returns the same result as JSON, and AI assistants that use the thinkhard.si MCP server get it as the check_answer tool. No key is needed.