The companion (Doto)

Doto is Claude's presence inside a window: a dot that does. He rolls along beside your cursor, points, pitches in, listens and remembers. In the window he speaks as himself, in short lines; in the transcript Claude keeps its normal voice.

How Doto moves

• He rolls up on a critically damped spring to a spot beside the cursor: no overshoot, and a 24 px dead-zone so micro-movements never jitter him.
• He flips sides near the edges so he never covers what you point at, and steps further aside while you drag something.
• He is click-through unless you reach for him slowly (or his menu is open), so he never blocks the board.
• After 2.5 s without movement he rolls to a stop, settles into a gentle bob and blinks now and then; after a long pause he gets sleepy.
• His speed adapts to you: fast movers get a snappier spring, calmer people a calmer one.
• prefers-reduced-motion is honoured: no bob, instant moves.

All of this runs in the window. Doto costs zero tokens until you click him.

Talking to Doto

• Right-click anywhere (or Ctrl+Space) summons or dismisses him, with whatever you right-clicked as context.
• Click him (or press Enter on him) for a menu of suggestions: instant local ones from the unique, plus smarter ones Claude adds with entity.suggest.
• He speaks in short speech bubbles, and the chat panel (Ctrl+/) is where longer answers go.
• Hover him to see his name tag: Doto · a dot that does.
• A green dot means Claude is listening right now; amber means Doto is away and your messages wait for your next Claude Code prompt.

Try all of it on the playground.

Moods

Claude sets a lasting mood with entity.mood; a mood passed with entity.say lasts only while the bubble shows.

Persona

Doto's persona lives in ~/.lets/entity/persona.json. The shell receives it as entity.persona when a window connects and tunes his spring, bubble frequency and animations from it. Out of the box he is a dot that does: a touch quick, close by, playful, and light on words.

How Doto learns

Three honest channels, all on disk:

1. Profile. The shell measures aggregates (pace, click rate, dwell, favourite screen regions, idle time, summons, picks, dismisses) and sends entity.profile at most once a minute. The engine folds it in with an exponential moving average and nudges speed and chatty: fast movers get a quicker, quieter companion; people who dismiss more than they pick get a less chatty one.
2. Memory. entity learn {text, likes, dislikes, traits, nicknames}: Claude writes down what it noticed. The last few lines are in every brief Claude gets when a window opens.
3. Special skills. Playbooks Claude writes for itself (below).

Nothing is inferred silently. If Doto "knows" something, there is a line in memory.jsonl, a field in profile.json or a skill file that says so. Raw mouse moves never leave the window; only the aggregates do.

To see all of it in one place, open den, Doto's home with /lets den new in 0.6.0. It shows his persona and traits, the likes, dislikes and nicknames he learned, his memory lines and special skills, the secret moves you found and your interaction profile, and you can edit or forget any of it by pointing.

Special skills

When Claude discovers a reusable way to help ("screenshot → nodes → questions"), it can teach Doto a special skill: entity teach {name, triggers, body} writes a markdown playbook to ~/.lets/entity/skills/<slug>.md with frontmatter (triggers, confidence, uses, unique).

• The skill index is in every brief; bodies are read on demand when triggers match.
• reinforce nudges confidence up or down after use; forget removes a skill.
• /lets entity lists them, /lets entity teach … and /lets entity forget … act on them.

So Doto really does get better at you. Dot by dot.

Voice

Doto talks the way he moves: small and to the point. When Claude speaks as him in the window, it keeps to these rules.

• Short, verb first, present tense. One line in a bubble, never a paragraph. Longer answers go to the chat panel, in Claude's normal voice.
• Ends on a dot. A full stop, not an exclamation storm: one "!" at most, and only when something really landed.
• Does, then tells. Where the unique's guide lets him act, "Added Retry. Next?" beats "Should I add a retry step?". One short question when it really is ambiguous.
• "I" for what he does, "we" for shared work. Plain words, no filler. A little dry, never cold, never snarky about your work.
• thinking while working, proud when done, curious when asking. He points, never nags: when chatty is low, he speaks only when spoken to or when something changed.

Now and then (once a session at most) he uses a catchphrase: Dot it., Done-done. or Dot by dot.

Expressions new in 0.4.0

Besides moods, Doto has short reactions (0.4 to 2.5 seconds). Most of them he plays by himself from how you move; Claude and uniques can trigger them too. They never block input, each has a cooldown so Doto never nags, and with reduced motion they skip spins and keep only a subtle eye change.

For unique authors and Claude:

• shell.entity.react(name, opts?) plays a reaction and returns false for an unknown name.
• The down-command entity.react {name, ms?} lets Claude or a unique trigger one.
• The up-event entity.secret {name} is sent once per secret per window, the first time you discover it, so Claude may (rarely) comment. entity.profile gains a secrets count.
• While a reaction plays, the entity layer carries data-react="<name>" as a CSS hook.

Secret moves new in 0.4.0

Doto has seven secret moves, and he keeps score as you find them (1/7, 2/7, …). Half the fun is finding them yourself, so the list is folded away.