Hire Your First Digital Employee

First decide what failed: the instruction or the knowledge. Ask the assistant where the answer came from. If it points at your file but read it wrongly, the problem is in the knowledge. If it points at nothing, the problem is in the instruction.

Put the fix where the cause is: into the instruction on the platform, or into the file you then re-upload. A correction said in the chat only lasts until that conversation ends.

If you are correcting the same thing a second time, it is not bad luck — it is a gap that keeps coming back. That sentence belongs in the instruction permanently.

Not the way a person does. Only two things persist between conversations: the instruction and the knowledge files. Anything you paste into a single conversation disappears with it.

Some platforms add their own "memory" about you. It is convenient, but it is not the place for facts that have to be right — keep those in files, where you know exactly what the assistant sees.

The practical rule: if something must still apply tomorrow, it has to be in the instruction or in a file. Not in a conversation.

An assistant prepares a result and hands it to you. You decide whether to use it. An agent takes actions on its own — it sends, saves, or changes something in another system.

The difference is not technical, it is about risk. A bad result from an assistant costs you a minute to fix. A bad action from an agent reaches your client or your data immediately.

That is why this goal builds an assistant. An agent is worth considering only once the assistant has done the same task for weeks and its results no longer surprise you.

You can, but it usually makes things worse. An instruction written for one task can be specific. An instruction written for five different tasks has to be general — and a general instruction gives general answers.

The same is true of knowledge: the files needed for client replies get in the way when the assistant is meant to summarize meeting notes.

A practical rule: one task, one recipient of the result, one format — one assistant. When any of those three changes, that is usually the sign for a separate assistant.

When the first one has survived two weeks of real use and you still open it. That is the only proof the mechanism works for you and not just in theory.

The second one is much faster to build: you already know the platform, you have a pattern for the instruction, and you know what a good knowledge pack looks like. It is usually a matter of fifteen minutes, not half an hour.

Three half-configured assistants are worth less than one you actually use.

If the change is small — a different answer format, a different tone, one more rule — just edit the instruction. It applies from the next conversation.

If the task itself changes, create a new assistant and keep the old one. Overwriting a working configuration is the most common way people lose something that was already serving them.

After any significant change, re-run your tests. An instruction that worked for the previous role does not have to work for the new one.

Not by the calendar, but by events. Update when something the assistant bases answers on changes: a price list, an offer, a procedure, contact details.

Beyond that, one review a quarter is enough to catch files that quietly went stale. Write that date into your starter kit, because otherwise nobody remembers it.

An out-of-date file is more dangerous than a missing one: the assistant will quote it with complete confidence.

Yes, and it is one of the reasons this goal has you write the instruction first and pick the platform second. The instruction and the files are yours — they work the same everywhere.

Moving usually means creating an assistant on the new platform, pasting the same instruction and uploading the same files. The conversation history stays in the old tool.

After moving, run both checks from module 5 — platforms treat files differently, so it is worth confirming the new assistant really reads them.

Anything you would not want to see in someone else's answer: clients' personal data, passwords, keys, account numbers, documents covered by a confidentiality agreement.

In most tasks they are not needed. An assistant meant to draft a client reply needs the price list and the rules — not a list of names.

If a file contains both, prepare a version without the sensitive parts. That is five minutes of work that removes the whole problem.

Usually the rule is too general. "Don't make things up" does not work. This does: "a service outside the price list always gets the answer: individual quote".

A rule works when it says what the assistant should do INSTEAD of guessing. A bare prohibition leaves it with no way out, so it picks the most plausible-sounding answer.

Also check whether the answer was in your files at all. If it was not, this is not invention despite the rules — it is a gap in the knowledge.

The practical limit is about five per task. Not because platforms will not accept more — they will. Because beyond that you stop knowing where an answer came from.

If you have ten candidates, ask about each one: if the assistant quoted this file word for word, would I be happy? Leave the ones you hesitate over out of the pack for now.

You can build an assistant with no files — it will run on the instruction alone. That works for tasks where the form and the way of working matter more than your specific data.

If the task needs your facts, the simplest first file is two or three examples of the work done well. That is enough for the assistant to learn your standard.

Knowledge can be added later. An assistant with no files is a better starting point than an assistant you never built.

For building and testing the assistant from this goal — usually yes. Free-plan limits are mostly about message counts and file sizes, not about whether you can create an assistant at all.

Limits change often, so check them on the platform's own page on the day you start. We do not promise that anything is free "forever".

Pay only once you hit a specific wall in daily use. Then you know what you are paying for.

That is what the starter kit is for: where the assistant lives, and the three commands you actually use. Three saved sentences make more difference than they sound like they would.

If you still come back after a month unsure where to start, open the assistant's instruction. It is written in your own words and will remind you what you built it for.

When you have run the same task with the same assistant for a few weeks and you accept the result almost without edits every time. Only then do you know what you would actually be automating.

Automation does not fix a weak instruction — it repeats it faster. Every mistake you catch by hand today will travel on by itself tomorrow.

Start with the most repetitive fragment, not the whole process. And keep a moment where you look at the result before it goes out.

In this task — no. The assistant prepares a draft; the decision and the responsibility stay with you. That is why this whole goal ends with testing, not with handing over the controls.

What does change is where your time goes. Instead of writing from scratch, you read and correct. On a task you do every week, that is usually the biggest saving in the whole process.

Then sharing stops being a nice-to-have and becomes a platform criterion — which is why it is one of the seven decision questions in module 4.

Before sharing, read the instruction once more through that person's eyes. Sentences obvious to you can be unreadable to someone without your context.

Agree who edits the instruction, too. An assistant with five users and no owner stops answering properly for anyone within a month.

Not from an impression, but from repetition. One good answer proves nothing — it is the fifth or tenth run that shows whether the result is predictable.

Concrete checks: does the answer hold the format, does it cite your files, does it ask instead of guessing, and how many of your edits does it need.

The simplest measure is a practical one: do you reach for it when nobody tells you to. If you do, it works.