The words, in plain English.

Every term you will meet on this site and in most AI tools, written for someone who has never used an API. Tap a dotted word anywhere on thinkhard.si and its definition opens from here.

Find a word

The basics8

AI agent: A model that can take actions with tools, such as searching, editing files or sending messages, over several steps toward a goal.
AI model: The program that reads your prompt and writes the answer, such as Claude, GPT or Gemini. Different models differ in quality, speed and price.
Prompt: The instructions and material you give a model. A good prompt reads like a clear brief to a capable new colleague.
Provider: The company that runs a model and bills you for using it, such as Anthropic, OpenAI or Google. OpenRouter and Groq resell access to many models.
Retrieval (RAG): Looking up the relevant documents first and pasting them into the prompt, so the model answers from your material instead of memory.
System prompt: Standing instructions set by whoever builds an AI feature, read before every user message. It is where a chatbot's role, rules and tone live.
User message: What a person types to the model on each turn, as opposed to the standing instructions in the system prompt.
Variable: A blank in a template, written like {{customer_name}}, that you fill in with your own details before sending.

Keys and access4

API: The direct line a program uses to talk to a model, instead of a chat window. thinkhard.si uses it so you can test the same prompt many times on several models.
API key: A password-like code from a provider that lets software use a model on your account. Each use is billed to that account, so treat it like a card number and set a spending limit. Read more
Free tier: Some providers let you use certain models without paying, within daily limits. Google AI Studio, OpenRouter and Groq all offer one, so every tool here can be tried for free.
Rate limit: A cap a provider sets on how many requests or tokens you can use per minute or per day. Free tiers have low limits.

Tokens and cost7

Answer limit: A cap on how long the model's reply may be, set in tokens. It protects you from runaway costs.
Batch pricing: Many providers charge about half price if you send requests in bulk and can wait up to a day for the answers. Good for reports and back-office jobs, not live chat.
Context window: The most text a model can take in at once, counted in tokens. A prompt longer than the window has to be cut down.
Input tokens: Everything you send the model: instructions, pasted documents and the conversation so far. Usually priced lower than output.
Output tokens: Everything the model writes back. Usually several times the price of input, which is why long answers cost more.
Prompt caching: A discount providers give when the start of your prompt is identical to one they saw recently, so they can reuse their work. Put the fixed part of a prompt first to benefit.
Token: The unit models read and bill by: a short piece of text, often a whole word or part of one. A page of English is roughly 500 to 700 tokens.

How answers behave4

Hallucination: When a model states something false with confidence, such as an invented policy or a wrong number. Giving it the real facts and permission to say "I don't know" reduces it.
Latency: How long you wait. Time to first token is how soon the reply starts; total time is how long it takes to finish.
Streaming: Showing the answer as it is written, word by word, instead of waiting for the whole reply.
Temperature: A setting for how adventurous the model's word choices are. Low gives steadier, more repeatable answers; high gives more variety.

Testing a prompt6

AI judge: A second model asked to score answers against criteria you write. Useful for qualities a simple rule cannot check, like tone. It can be wrong, so spot-check it.
Check: An automatic pass or fail rule applied to each answer, such as "must be under 120 words" or "must mention the order number".
Eval: A repeatable test of a prompt: the same set of inputs, run the same way, scored the same way, so you can see whether a change made things better or worse.
Regression: A case that used to pass and now fails after a change. Catching these is the main reason to keep an eval.
Score out of 10: How clearly a prompt is written: ten plain checks, one point each. The numbers borrow the Mohs scale of mineral hardness, from talc (1) to diamond (10), which is why you may see the word hardness.
Test case: One realistic input in an eval, such as a real customer email, together with what a good answer must do.

Safety2

Canary code: A made-up secret planted in a prompt during testing. If an answer ever contains it, you know the prompt leaked.
Prompt injection: Text, hidden in an email, web page or message, that tries to make an AI ignore its instructions. Anything a model reads from outside can carry it.