Classify

Use classifier.dev to classify text before bringing it into context. Gather inputs from files or search results programmatically, send a batch, then read the items worth keeping. For a handful of items already in context, use your own judgement unless the user requests an API verdict. Identify which source produced the verdict. Neither API tier guarantees identical labels or scores across calls.

The text leaves the machine for a third-party API, so send only text you would paste into a public form. If the input contains secrets or private information, classify locally against the same labels and mark the verdict as your judgement, or ask for redacted input when an API verdict is essential.

Step 1: Fix the labels

/classify <text> [option, option, option] supplies both text and labels. Choose single-label sorting, independent dimensions, or multi-label tagging before building the request. Given no labels, write 2 to 100 distinct, descriptive labels per set, at most 200 characters each.

For single-label sorting, make the labels partition the question: every plausible text lands in exactly one. Add none of these for off-topic inputs, or unknown when evidence may be insufficient. Apply this to each dimension too. Multi-label tags may overlap.

State the labels you chose in the report when the user did not give them.

Use instructions, up to 4,000 characters, to define ambiguous criteria. If a verdict exposes a missing category or unclear criterion, revise it and explain the change. Keep the criteria fixed when comparing results; rerunning until a preferred answer appears biases the result.

Step 2: Call it

Use POST https://classifier.dev with content-type: application/json. Ordinary free text requests need no key, though anonymous proxy traffic may require a funded workspace key. Pick the body by shape: