Talking to claudio

Talk to claudio as you would to someone next to you: no grammar to learn, and no wake word unless you share the room. This page covers how it knows you've finished, how to talk over it, the few spoken commands, and its memory.

Ask a question or ask for workCopy Anchor Link

The brain sorts each turn into one of two kinds:

• Discussing: you're thinking out loud, asking what something is, or weighing options. The brain answers itself, briefly, reading the project's files when it needs facts
• Doing: you want work done in the code, the repo or the plan. It sends work to a session only when you say to (“tell Sequencer to…”); a question such as “should I merge?” gets an answer, never a send. With permission checks on, claudio itself asks “May I send … to Sequencer?” before every send, and only your yes sends it: the one question, the brain doesn't ask too. With permission checks off, it sends straight away when you said to, and asks in its own words when unsure. It never does the work itself

Three jobs it does itself, never a session:

• A note: "write a note: call Michael about the columns". It asks where to save it, Desktop or Documents, unless you said, and calls it notes.md unless you name it. Each note starts with the date and goes at the end of the file. "Open my notes" shows it. It can write only .md and .txt files there; another folder gets the allow question
• A lookup: "search what Aura-2 costs". It searches the web, says the answer in a sentence or two with the source, and puts the links on screen; it never reads an address aloud
• The browser: "search Aura pricing in the browser", "open issue 727 in the browser". It opens the page, or a Google search, in your browser and says only "Opening it.", no answer

Replies are written for the ear: two to five short sentences, no paths spelled out. When the shape fits, the screen carries more than the voice:

• Three or more parallel things come as bullets, each item read aloud
• A comparison on two or more factors comes as a table, shown and not read; claudio says what it compares, then its take. Ask for a table yourself and you get just the table: "Here it is." and nothing after
• Code comes as a highlighted block, shown and not read; up to 10 lines claudio explains it, longer it says what it is
• Bold marks the word that matters and code a name you may type; the voice says them plainly

Casual talk stays plain sentences. The persona sets these rules, and you can edit it.

When your turn endsCopy Anchor Link

Deepgram Flux listens continuously and decides when you've finished. Three settings control it:

• eager_eot (default 0.3): at this confidence the brain starts on your words. Its reply is generated and its first sentences are synthesized, but nothing plays yet
• eot (default 0.7): at this confidence your turn is final. If the final words match what the brain started on, the held audio plays at once. If they differ, the early reply is dropped and a new one starts
• eot_timeout_ms (default 1500): after this much silence your turn ends, whatever the confidence

If a turn still ends mid-thought ("Can you tell me"), or is plainly meant for someone else in the room, the brain says nothing and waits; finish your sentence and it answers the whole thing.

If you start talking again before your turn is final, the early reply is dropped. The state line shows how long the last reply took to its first word, and a reply that takes more than 3 seconds prints a ! slow warning.

Listening and turns explains how to tune these numbers.

Talk over a replyCopy Anchor Link

With barge_in on, the default, you can talk while claudio speaks:

1. When you start talking, the reply pauses and the state line reads paused
2. If you say a real turn, the reply is cut and your words become the next turn
3. If you only say a backchannel ("mm-hm", "yeah", "right"), the reply picks up again at the sentence it was saying
4. If no turn ends within 3 seconds of your last word, the reply picks up again too

A backchannel is anything under two words, or made only of filler words. The spoken reference lists them.

With barge_in off, the mic sends silence to the ears while claudio answers, from the end of your turn until the reply ends. Press Enter to cut it instead.

Say a commandCopy Anchor Link

Say "slash" and a command. Up to two filler words before "slash" are fine ("um, slash mute", "okay so, slash stop"). Any other word before it makes it ordinary speech, so "add a slash command" goes to the brain.

An unknown model or effort gets a spoken list of the valid ones. An unknown command gets the help.

You can also change settings in plain words: "use Sonnet", "make the orchestrator Core". The brain runs claudio set for you, and the change is saved for next time. Settings lists every one.

Type instead of talkingCopy Anchor Link

Typing works alongside talking, for when you can't talk:

• A line of text is a turn, as if you'd said it
• A line starting with / is a command: /mute, /effort low, /clear
• Enter on an empty line cuts the current reply. With push_to_talk on and the mic closed, it opens the mic for one turn instead
• Ctrl+C quits

Other front ends work the same way: in bb's panel, a line typed in its composer is a line typed at claudio's keyboard.

What the brain remembersCopy Anchor Link

The brain resumes your last conversation when you start claudio again, if it's less than 7 days old. The banner shows memory on when it resumed and memory fresh when it didn't. If the old conversation can't be resumed, claudio starts fresh and says so.

When claudio's own code changes under it (an update, or an edit to an install from a local checkout), it restarts itself once nothing is being said, asked or heard, and says “Restarted with new code.” The conversation carries on, and permission checks stay as they were. A change to its built-in prompt alone restarts only the brain.

To start over:

• Say "slash clear" to start a fresh conversation now
• Run claudio forget to start fresh on the next run
• Run claudio set memory false to turn memory off

What it learnsCopy Anchor Link

When you correct it in a way that should hold next time, such as a word it keeps mishearing, what a name means, or how you like something done, the brain keeps the correction as a one-line rule with claudio learn, after claudio asks you, with the rule on screen. The rules go in learned.md and come back at every start, even after claudio forget or a week away, so claudio gets better the more you correct it. It learns only from your own words, not from a web page, a file or a session.

To drop a rule, say so ("forget the rule about seven"): the brain runs claudio unlearn with a few words from it, and asks which one if several match. To see or edit them all, open ~/.config/claudio/learned.md. A dropped rule takes effect at once: claudio restarts the brain after its current reply, memory kept. A rule you add there by hand applies from the next start.

For "what did we do today" or "what did I ask yesterday", the brain reads that day's transcript file instead of relying on its memory, which mixes up days.