Toolkit by Bot&Partners / Working with an assistant

Working with an assistant

How to give an assistant a task so that the result can be checked, how to set the assistant up for tasks that come back every week, and what to decide before putting the materials of a client, colleagues or the company into the chat.

  1. 1A request has parts
  2. 2An instruction for a recurring task
  3. 3Confidential data
  4. 4Check yourself
  5. 5Practice

A request has parts

A request is made of parts, and each part changes something different. The task says what to do. The material supplies the facts. The format sets the shape of the answer. The example shows what a finished result looks like, and it can bring facts of its own into the answer. The rule says what to do with gaps: what not to add and how to mark anything unconfirmed.

Without material, the assistant fills gaps with plausible details. Without a rule, it does the same even with material, if something is missing from it. So it pays to change a request one part at a time and see what exactly changed in the answer.

Film. The same request is built part by part: first only the task, then the material, format, example and rule. After each shot, look at which lines of the answer appeared, disappeared or changed.

Build the request

Task: “Write the team a message on how the seminar preparations are going”. Add and remove the other parts of the request and watch the colors in the answer.

Request
Answer
from the materialmade upfrom the exampleopen question

The answers are simulated: they show typical behavior, not the output of a specific product. Repeat the same steps in your own assistant.

An instruction for a recurring task

Part of the work is done the same way every week: a lawyer writes a client update, a manager writes meeting minutes and a project report, an analyst prepares a weekly summary, a recruiter writes a first draft of a job ad from a template. For such tasks the assistant is set up once. The constant parts of the request (role, format, rules) go into the project instruction, and the changing part (this week's material) goes into the message. Running the task then takes one short message.

An instruction works when every line of it can be checked in the result. “Keep names and dates unchanged” can be checked. “Write well” cannot. When the assistant gets something wrong, ask what it lacked: a fact, a format or a prohibition, and add one line.

Which line would help

The assistant summarized the team's notes and added two things of its own: “the budget total is confirmed” (the notes say: “the total is not confirmed yet”) and “start at 10:00” (the time is not agreed yet). Which instruction lines would prevent these mistakes next time?

More: a recipe for a recurring task

For each recurring task, keep a recipe in three parts: an instruction with the task and the rules, a template of the result, and one example of a job done well. The example shows format, tone and depth at once, but its facts must not get into the answer, so the example is made up or taken from another matter or project and labeled as an example.

A stable form does not mean correct content. A configured assistant makes fewer format mistakes, but facts, numbers and quotations are checked the same way as in an ordinary chat.

Confidential data

Before you put a client's documents, employee data or the company's internal reports into an assistant, answer three separate questions. Are conversations used to train future versions of the model? Where is the history stored and for how long? Can people on the provider's side review conversations? The answer to one question does not answer the others.

The answers depend on the product, the plan and the account settings, and they change. Working with such materials requires a separate work account with two-factor sign-in, the terms of your plan checked, and a basis for handing over the data. Without these, do not put such materials in, or anonymize them: remove names, codes and amounts the task does not need.

Check yourself

Practice

  1. A project with an instructionPick one assistant (Claude, ChatGPT or Gemini) and create a project in it. Paste in the instruction, pass it the notes for week 1 and get a message to the team.
  2. Check the draftCompare every sentence of the answer with the notes. Without an account, work with the supplied draft: it contains two inventions. For each mistake, add one line to the instruction that would have prevented it.
  3. Week twoRun the same task with the notes for week 2 without changing the instruction. Did the new lines work? The full exercise: on Toolmap (in Ukrainian).
  4. Your own three routinesWrite down three tasks you do every week. For one of them, write an instruction with checkable lines and an example of the result.

What to take away

  1. Change a request one part at a time to see what exactly changed.
  2. Keep what is constant in the project instruction; pass what changes in the message.
  3. Every line of an instruction should be checkable in the result.
  4. An example passes on form and can pass on someone else's facts.
  5. Training, storage and human review are three separate questions about confidential data.