Teach mode: show your agent a workflow once

By Arjun Mehta · Updated 2026-07-05

Three ways BackBrain learns from you

You don't have to script or configure anything to teach BackBrain how you want a task done. There are three conversational commands that cover the full range of teaching situations: showing it a workflow, telling it a rule, and correcting it during a run.

Show It: the watch-and-learn mode

Say "watch me do this" at the start of a task. BackBrain enters observation mode: it records your actions without taking any of its own, following along as you navigate pages, fill fields, and click buttons. When you finish the demonstration, BackBrain plays back a plain-language summary of what it understood — the sequence of steps and the logic it inferred from your choices.

The summary is a checkpoint, not an automatic save. If the summary is accurate, you reply "save" and the demonstration becomes a stored procedure. If it missed something or misread a step, you can correct it before saving. Nothing is persisted without your explicit "save."

Tell It: persistent chat instructions

For rules that apply across many tasks on the same site, chat commands are more efficient than demonstration. Say "remember: always download in PDF format on this portal" and BackBrain stores it as a durable instruction, surfaced automatically every time it works on that site.

Forgetting is just as simple: "forget what you learned about that portal" causes BackBrain to list every matching memory and delete only after you type a confirmation. You are always shown exactly what will be removed before anything is gone.

Correct It: mid-run corrections

When BackBrain takes a wrong step during a live task, you can interrupt in plain language: "no, use the CSV export, not PDF." BackBrain stops, applies the correction, and proceeds. The correction is automatically saved — not just for this run but as a standing preference for future runs on the same site.

This makes mistakes self-correcting rather than recurring: the first time BackBrain gets something wrong and you fix it, it stays fixed.

What is and is not recorded

BackBrain is deliberate about what it does and does not memorise from a demonstration. It records which field you filled and which button you clicked — the structural actions. It never records what you typed into a field. Your password, your account number, the amount you entered — none of that is captured in the procedure.

When the procedure replays and needs a credential, it retrieves it from BackBrain's built-in encrypted vault at the moment of use, exactly as if BackBrain were running the task from scratch with full AI reasoning.

When to use each mode

• Show It is best for multi-step workflows with an unusual sequence that is hard to describe — portal navigation with non-obvious menus, or a multi-app workflow across two programs.
• Tell It is best for standing preferences: format choices, file-naming conventions, which contact to use for a recurring task.
• Correct It mid-run is best when BackBrain is nearly right but makes a consistent small mistake — one correction during the run fixes it for every future run.

Teach mode also activates automatically during a provider outage: if BackBrain stalls mid-task and you take over manually, it switches into watch-and-learn mode and records your steps. Read more in AI provider outage resilience.