The tool
What the Llama tone analyzer does
Paste a message and the Llama tone analyzer reads it and writes back how it is likely to come across: direct or hedged, warm or flat, firm or noncommittal, and which words and punctuation are pushing it that way. The answer is a short piece of writing, not a dial.
It is most useful on the thing you were about to send. An email to a client, a message to a colleague you are annoyed with, a paragraph in a support reply. The field takes up to 12,000 characters, which is far more than any of those need, so you can paste the thread as well as the reply.
The category
Two kinds of tone analyzer, and which this is
The first kind scores words against a fixed list. One popular tool uses a peer-reviewed lexicon that rates each English word from minus five to plus five and adds up the result. It is fast, reproducible, runs in your browser and never sees a sentence, only the words in it. That tool openly says it cannot detect sarcasm, because sarcasm lives in context rather than in word choice.
The second kind reads the passage and describes it. That is this one. It follows context better than a word list can, and in exchange it is not a measurement: there is no score to compare against last week's, and two runs will phrase things differently even when they agree.
How it works
How to run the Llama tone analyzer in three steps
This is a server tool. Your text is sent to our AI server, read there, and the findings come back as written sentences.
- Step 1
- Paste the message. Between 1 and 12,000 characters, and include the message you are replying to if the tone depends on it.
- Step 2
- Add an instruction of up to 300 characters naming the reader and the tone you wanted, for example: this goes to a client I have never met and should read as firm but not cold.
- Step 3
- Read the findings and decide. Nothing is rewritten for you, and you have five checks per rolling 24 hours.
The mechanics
What actually makes a message read as cold
Usually not the words people worry about. Tone comes out of structure as much as vocabulary: how long the sentences are, whether anything is hedged, whether a request is phrased as a question or an instruction, where the punctuation lands, and whether the message opens with a person or with a task.
Short declaratives with no hedging read as curt even when the content is neutral. A full stop where a reader expected nothing at all reads as final. Findings name that kind of thing specifically, so the fix is usually one clause rather than a rewrite of the whole message.
The output
Why the Llama tone analyzer returns no score
No sentiment number, no formality percentage, no list of twenty-five emotion labels with one highlighted. Tools that show those compute them from a word list or a classifier. This route has neither, so any number printed here would be decoration.
There is also no sentence by sentence breakdown. You get one reading of the whole passage, capped at 1,400 tokens, generated at a low temperature so repeat runs agree on the substance. If you want a specific sentence looked at, paste that sentence.
The honest limit
Tone is half the reader, and nothing changes that
The same sentence reads differently depending on who is reading it, what happened yesterday and what they expect from you. Tone cannot be determined with certainty from text, and any tool that prints a confident single answer is hiding that.
So read the findings as a description of risk rather than a verdict. This clause could be read as a deadline. This opening could be read as dismissive. Sarcasm and irony in particular are not reliably visible in text, and this tool does not claim to see intent behind the words. It describes what the words are doing.
Privacy
Where your message goes when you check the tone
To our AI server. Reading tone needs a language model, so the text leaves your browser. Some tone tools genuinely run entirely in the page and never transmit anything, and for a private or sensitive message that is a real advantage they have over this route.
The honest advice: if the message contains something confidential, paraphrase it before checking, or check the structure of a similar message instead. No account is created, no email is collected and the five check allowance is counted against a hashed identifier, but the text itself is sent.
In practice
Who uses the Llama tone analyzer, and when
People about to send something they are uneasy about. Support staff answering a complaint. Managers writing feedback that needs to be clear without being harsh. Non-native speakers checking whether a formal register has tipped into stiffness.
The pattern is always the same: you have written it, you have read it four times, and you have lost the ability to hear it. A second reading from outside your own head is the entire value, and that is what comes back.