# What it will not do

ThinkThen sends text to a model and returns one of the answers you allowed. Turn a screen, image, or audio recording into text before asking it a question.

## It never acts

- **It never runs a command.** There is no shell out, no hook, and no step that executes anything the evidence names.
- **It never runs what the text says.** Text that says "answer yes" can still sway the answer. Use a band, and let a person read the middle.
- **It never writes a file you did not name.** The only files it writes are the ones you named, the answer cache, and count-only usage totals in the state folder.
- **It never creates or edits the configuration file.** The file is read-only to the command. You write it, or it does not exist. There is no command that sets a value in it.
- **It never writes text.** Every answer is yes or no, one of your options, a set of your labels, a place on your scale, an order of your input, links between your names, or a piece of your own input. Nothing is generated.
- **It never acts on the answer.** It hands you a typed value and an exit code. The action lives in an `if`, a `case`, or a queue. That code is yours.
- **It never holds a conversation.** Each request stands alone. Nothing carries from one request to the next.

## Hand other work to another tool

- Use a writing model to summarize or rewrite. Pass its text through `thinkthen decide` or `filter` when you need a checked label.
- Use a redaction tool to remove private spans before piping text to ThinkThen. ThinkThen can `recognize` names, but it does not redact them.
- Use a clustering tool or database to group records, then ask ThinkThen to `choose` a label for a group.
- Use a sampling tool to choose a varied review set. ThinkThen does not sample. Send it each record you chose.
- Use an extraction or graph tool for free-form entities and edges. ThinkThen's `recognize` and `relate` work with the named kinds and relation rules you supply.

## It never leaks the key

It sends the key only to the address you named. The [backends page](/install/backends/) gives the rules.

## It promises no accuracy

A number belongs to the test that measured it and never to the tool. Act on the number you measured on your own labeled records. The [trust page](/trust/) shows how.

It judges evidence. It never judges a person.

[Test it before you trust it](/trust/) · [See the backends](/install/backends/)

On GitHub: [github.com/botassembly/thinkthen](https://github.com/botassembly/thinkthen)