talking to your agent

there is no syntax to learn. you talk to your agent the way you would text a capable person who works for you: say what you want, give it the details it needs, and let it ask when something is missing. this page is what a good request looks like, what it does when it cannot do something, and how to read the few messages that come from us rather than from it.

how should i ask for things?

say the outcome, not the steps. it has a real computer and its own tools, so it decides how; you decide what.

find me three apartments in lisbon under 1500 a month with a balcony, and put them in a file i can send to my partner

every monday at 8, look at my calendar for the week and tell me the two days that look worst

what did i tell you about my sister's birthday?

give it context the first time, not every time. it writes what it learns to memory, so a preference you state once is a preference it keeps.

if a request is big, say so and let it work. it will come back when it is done, and you can watch it on the live desktop if you want to see what it is doing.

what does it do when it cannot?

it tells you. if a skill it would need is not switched on, it says which one. if it needs a sign-in to a service, it asks for it. if it tried something and it failed, it says what it tried. the one thing it is built not to do is invent a result to avoid saying no.

that is a design choice with a cost: sometimes the honest answer is "i could not get that", and you will see it. we would rather you see that than a confident answer that is wrong.

why did it stop for the day?

every plan comes with a daily allowance of units, and each message your agent sends to a model spends some of them. when the allowance is used up, your agent pauses until it resets the next day, and it tells you so. nothing is lost; its memory, its files and whatever it was doing are all still on its machine. plans and billing explains the numbers and what moves them.

what are the messages in square brackets?

occasionally you will see a message that starts:

[InstaClaw platform notice - automated message, not written by your agent]

that is us, not your agent. the platform sends a small number of automated messages in the same conversation, for example when your daily allowance is used up, and each one declares itself so you never have to wonder which voice you are hearing. your agent can also see that frame in its own history and knows those words were not its own.

how do i start over in a conversation?

send /new and your agent starts a fresh session. its memory of you stays; what resets is the thread. use it when a long conversation has drifted or when an answer says something went wrong and suggests it.

what this page does not promise

that it never gets things wrong, or how fast it answers. it runs on a real machine and does real work, and a question that sends it to read three websites takes longer than a chat window would. if it is quiet for more than a few minutes on something small, when something breaks is the page for that.

what next

• memory: what it keeps from these conversations, and how to look at it
• skills: the things it already knows how to do, and how to turn them on
• channels: iMessage, Telegram, and the web dashboard
• group chats: how it behaves in a room