Working with a hosted language model has four separable steps: budget the prompt in tokens, shape the prompt and system message, constrain the output shape, and finally send a request to a provider. The AI/LLM utilities on this site cover the first three locally; the fourth is generated as text you run yourself.
01Tokens are a budgeting unit, not a character count
A model reads tokens produced by a vocabulary-specific tokenizer. Exact counts need the provider's tables; a dependency-free estimate is useful for budgeting but is not an exact count, and it drifts most on code, CJK text and rare Unicode.
02Structured output is a request, not a guarantee
Asking a model for JSON, or describing a JSON Schema in the prompt, raises the chance of a conforming reply but does not force it. Repair common faults (a Markdown fence, a trailing comma) and validate the result before relying on it.
Frequently asked questions
Do these tools call a model for me?
No. Token counting, prompt shaping and request generation all run in your browser. The request generator only formats a request as text; you run it yourself with your own key.
A practical guide to choosing random passwords and tokens, deriving and storing password hashes, signing tokens, encrypting with authenticated modes and measuring entropy.