Working with the Assistant

The lab builder — the chat panel beside the lab editor — is how you talk to the AI while it works on a lab. Beyond approving each step of the plan, you can guide the assistant in plain language, answer its questions, review the files it changes, and come back later to edit a finished lab. This page covers those interactions.

Where the assistant appears

The assistant shows up in one of two places, depending on the lab:

On the assistant page — while a lab is being built from scratch with AI and its plan isn’t finished yet. The chat sits on the left with the plan beside it, and the header’s Assistant button takes you there.
Docked beside the editor — for a finished lab, or any lab that was never AI-generated. The Assistant button becomes a toggle that pops the chat out to the left of whatever page you’re editing, so you can reach it from anywhere in the lab.

The chat

The chat streams what the assistant is doing as it works and is where you give it direction. Type into the message box — for example “add a short quiz at the end of chapter 2” or “make the first chapter more hands-on” — to steer it.

While the assistant works, an activity line at the bottom of the conversation shows what it’s doing and how long the turn has been running — and, when it’s working through a task list, the task it’s on above the step it’s taking.

You steer the assistant between turns, not during them: the message box accepts your next instruction once the turn ends. Turns are short and each one reports what it did, so if it takes a wrong direction you can correct it on the next message rather than catching it mid-write.

Answering questions and approving changes

Sometimes the assistant needs input before it can continue. These prompts appear just above the message box:

Questions — when a request is ambiguous, the assistant may ask a clarifying question. Answer it to let it proceed.
File change approvals — before making certain edits, the assistant may ask you to approve a change to a specific file. You can accept it, or deny it and add feedback explaining what to do differently.

Approving and denying are how you keep the assistant on track — think of them as steering the conversation rather than a security gate.

Stopping and restarting the session

The session menu beside the chat title has two session-level actions:

Stop session — ends the run and leaves the assistant idle. Your next message starts a new session.
Restart session — ends the run and immediately starts a fresh one.

Both are confirmed first, because the assistant keeps no memory across sessions: the conversation goes with it, while everything already written to the lab is kept.

Use them when a session has wandered off, or when it’s working from context that’s no longer current — for example after you discarded the changes it was building on. They act on the session as a whole rather than the turn in flight, so reach for them between turns.

Reviewing generated files

On the assistant page, switch the right pane from Plan to Files to browse everything the assistant has generated — the lab’s content, activities, sandbox configuration, and other files. With the docked chat you have the whole lab editor beside you instead, so you can watch its changes land on the page you’re working on.

Your edits and the assistant’s edits accumulate together on the same working branch, so you can freely mix chat-driven changes with manual ones. Publish, Review changes, and Discard are disabled while a turn is running — using them mid-write would fork the session — and become available again as soon as it ends. Once the assistant has written something they light up, even if you haven’t touched the lab yourself. As with any lab, you review the combined changes and save or publish them through version control.

Editing an existing lab

Generation isn’t a one-way door. Open a finished lab, toggle the Assistant panel, and ask for what you want changed — shorten a chapter, update the sandbox, add a new quiz. The assistant reads the lab as it stands, loads your company and product research for context, and makes the change directly.

There’s no plan to approve here, and none is reconstructed for you: the chat is free-form and makes a best-effort edit to the built lab, touching only what you asked about. You can also just ask questions — “what does chapter 2 cover?”, “which resources does this lab use?” — and it answers from the lab without changing anything.

The assistant starts idle on a lab like this and opens a session when you send your first message. Its edits land on the same working branch as your manual ones, so nothing reaches users until you publish.

Keeping an eye on usage

The assistant’s work is billed to your team’s own Anthropic API key, so your spend is visible in the Anthropic console rather than in Instruqt — see Usage & limits for what’s being measured, and Using your own API key for what happens when a key fails.

Recovering from problems

If the assistant runs into a problem — for example it fails to start — the chat shows an error and a Try again action to provision a fresh session; sending a message does the same thing.

If the connection drops mid-turn, the chat reconnects on its own and replays the conversation, so you don’t need to refresh. Refreshing is also safe: the session reconnects and picks up where it left off, so you won’t lose approved work.