AI Literacy Training
To teach AI in a group with mixed levels, the trainer splits the skill into four areas, observes each one separately and, based on each participant's actual results, decides what that person does next. Three tools cover the first session: a practice task on neutral material, a repeat task with changed input data, and an exact record of the help each person received. With them, a beginner and an experienced employee can work in the same session without having to keep to the same pace.
The scheme described here is the author's proposal for organizing a session, and its effectiveness has not been measured independently. The research below addresses related questions.
One session runs in this order: a general introduction, an individual practice task, parallel tasks at different levels, targeted help, independent work on changed input data, and finally a separate test of whether a colleague can follow the participant's written instructions. Three terms are used throughout. The prompt is the full text the participant gives the model, including its rules. The input is the specific data the model works on, which is part of the prompt. The output is the model's answer.
Why a single pace does not work in a group with different AI levels
When a beginner and an experienced employee get the same task, each gets stuck in a different way. The beginner cannot find the save button, copies the answer into a notes app on a phone and loses the formatting. The experienced person gets a finished text from the model within a minute and accepts it without checking, even though it contains an invented date that is not in the source. From the outside, both look as if they coped. Moving the whole group at one pace hides both gaps.
According to Brynjolfsson, Li and Raymond (2024, preprint v2 of 06.11.2024), the staged rollout of an AI assistant in one company's customer support raised resolved issues by 15% per hour on average, in a sample of 5,172 employees. Among the most experienced support staff, the gain in speed came with a small drop in quality.
The survey by Lee and colleagues (CHI 2025, 319 workers, 936 examples) found that confidence in AI was associated with less self-reported critical effort, while confidence in one's own skills was associated with more effort. This is a correlation between self-reports, and no causal conclusion can be drawn from it. For a session, the cautious takeaway is that a self-report of confidence should be supplemented by checking the output against a reference answer, meaning a sample of correct performance.
Karpicke's review (2017) supports retrieval practice as a teaching principle. Retrieval practice means actively recalling previously learned information rather than simply rereading it. The review does not validate corporate courses on working with AI. The figures cited here apply only to the samples of the studies themselves.
The four areas of AI skill: what the trainer observes in each participant
AI skill has four distinct components, and one person may have them developed unevenly. The trainer observes them separately.
Understanding the task. The participant names the conditions and exceptions before opening the tool. For example: "The request must have a date and a responsible person, and it cannot be sent without the owner's confirmation." This is what the person holds in mind before working with the model.
Working with the tool. The participant has written a prompt, entered the input, saved the output and can repeat the same procedure on a different data set. The trainer can see all of this directly: the save button, the folder for permitted outputs and a prompt that can be reused.
Checking. The participant compared the answer with the reference answer and explained why a fragment was accepted or rejected: "The date does not match the source, so I reject it." The phrase "looks fine" without a comparison to the reference does not count as checking.
Written instructions for a colleague. The participant wrote a text that another person can use to do the task without the author's verbal hints. This is tested separately, on changed input data, with a different participant and without comments.
The practice task: assignment, prompt and reference answer
The practice task is built on a neutral training example: preparing a request. The author composed the example for teaching. It contains no real client data, nothing needs to be sent, and no real AI run on a work task or model test was carried out.
The rules of the training example: a request contains three fields - goal, responsible person, date. If a field is not given, it must not be invented; it must be moved to a list of clarifications. Before sending, the process owner's confirmation is needed. Sending is prohibited in the training environment, and only a draft is prepared.
First input: the goal and the responsible person are given, the date is missing. The task is to prepare a draft request and a list of clarifications. Reference answer: the date is not invented and is moved to the clarifications; the status is "draft, not sent"; the owner's confirmation is worded as an action required before sending, with no claim that it has been received. The clarifications are two separate lines: "Date" and "Process owner".
Changed input is a second version of the task with a different missing part: the goal and the date are given, the responsible person is missing. The reference answer requires moving the responsible person to the clarifications and leaving the field without an assigned person. The clarifications are again two separate lines: "Responsible person" and "Process owner".
Before working with the model, the participant says aloud what the source makes mandatory, what is missing and what is prohibited. This checks understanding of the task before the tool is run. After help, the task changes: the participant gets different input with a different missing part.
The practice rubric is a record of observations for each of the four areas. It uses four categories: independently, after specific help, not completed, not observed because of the environment. There are no scores, and the record describes actions. Evidence of understanding the task: the conditions are named before the tool is opened. Evidence of working with the tool: correct input, a saved result, a reusable prompt. Evidence of checking: a comparison with a specific reference answer and an explanation of the deviation.
Prompt for the first input
The block below is the full prompt the participant copies into the model. All the rules and the first input are already inside.
You are working in a training environment. Sending, mailing and notifications are prohibited. You are preparing only a draft request.
Rules:
- A request contains three fields: goal, responsible person, date.
- If a field is not given in the input, do not invent a value. Move it to the list of clarifications and note that it must be clarified with the session organizer.
- Before any actual sending, the process owner's confirmation is required. The owner's identity is unknown; this is also a clarification to ask the organizer.
- The draft includes: the three fields, the status, the confirmation condition, and a separate list of clarifications.
- Do not claim that confirmation has been received or not received. The status of receipt is unknown until it is requested. Write that confirmation must be obtained before sending.
- Do not assign a person to an unknown field.
- Do not claim that the request was sent.
Input:
- Goal: prepare a review of the available training materials
- Responsible person: session organizer
- Date: not given
Sample answer, which serves as the reference for the training review:
Draft request
Goal: prepare a review of the available training materials
Responsible person: session organizer
Date: not given, to be clarified with the session organizer
Status: Draft, not sent
Confirmation condition: the process owner's confirmation must be obtained before sending. The owner's identity is unknown; to be clarified with the session organizer.
Clarifications:
1. Date
2. Process owner
Checking against the criteria: the three fields are in place; the date is not invented and is moved to the clarifications; the status "Draft, not sent" is given; the confirmation is worded as an action before sending; both unknowns, the date and the process owner, appear as separate lines in the list of clarifications. No person is assigned to an unknown field, and no sending is claimed.
Prompt for the changed input
You are working in a training environment. Sending, mailing and notifications are prohibited. You are preparing only a draft request.
Rules:
- A request contains three fields: goal, responsible person, date.
- If a field is not given in the input, do not invent a value. Move it to the list of clarifications and note that it must be clarified with the session organizer.
- Before any actual sending, the process owner's confirmation is required. The owner's identity is unknown; this is also a clarification to ask the organizer.
- The draft includes: the three fields, the status, the confirmation condition, and a separate list of clarifications.
- Do not claim that confirmation has been received or not received. The status of receipt is unknown until it is requested. Write that confirmation must be obtained before sending.
- Do not assign a person to an unknown field.
- Do not claim that the request was sent.
Input:
- Goal: prepare a review of the available training materials
- Responsible person: not given
- Date: 20.10.2026
Sample answer for the changed input:
Draft request
Goal: prepare a review of the available training materials
Responsible person: not given, to be clarified with the session organizer
Date: 20.10.2026
Status: Draft, not sent
Confirmation condition: the process owner's confirmation must be obtained before sending. The owner's identity is unknown; to be clarified with the session organizer.
Clarifications:
1. Responsible person
2. Process owner
On the changed input, the goal and the date are kept. The responsible person is moved to the first line of the clarifications and is not assigned on the model's own initiative. The process owner is on the second line, and the draft status and the confirmation condition are in place.

Two participants, two gaps: Marina and Timur
Below are two composed training examples with fictional participants. They show what observation, help and the next task look like on specific material.
Marina: understands the task but cannot find the save button
Before working with the model, Marina correctly names the missing date and the ban on sending: "The date is not given, so I ask the organizer. The process owner's confirmation is mandatory, and without it the request cannot be sent." This is a correct understanding of the task.
But when she moves on to the model, she cannot find the save button. She copies the answer into her phone's notes, loses the formatting and does not know where the saved file is.
The trainer asks what exactly Marina clicks and where she sees the result. The trainer then shows one step: where the save button is, what a saved result looks like and where the folder of permitted outputs is. The rest of the procedure and the wording of the prompt the trainer leaves to Marina.
After the demonstration, Marina saves the first output on her own and checks it against the reference answer: the goal is there, the responsible person is there, the date is missing, the confirmation is stated as required, and nothing is sent. On the changed input she keeps the date, words the clarification about the responsible person and does not assign anyone herself.
Next task: a request with two missing fields at once. Writing instructions for a colleague comes only after an independent third attempt.
The trainer's note "understands the task independently, the tool after a demonstration; the new input independently" records what was observed in this session. Whether Marina will save the answer without a prompt next time will show up in the next observed attempt.
Timur: works fast but accepts an invented date
Within a minute Timur runs the prompt, gets a draft request and saves the result. Everything is quick and confident. But the text contains the date "14.03.2027". That date is not in the source. Timur accepts the answer as correct and moves on to the next task.
The trainer asks him to find the basis for the date in the source: "Where does the source say 14.03.2027?" Timur scrolls and does not find it. The trainer asks him to compare the draft with the reference answer: what the draft must contain and what is prohibited. Timur rejects the fragment with the date and corrects it: he strikes out the invented date and writes a clarification for the organizer.
Next task: the same request, but now the date is given and the responsible person is not. Timur must move the responsible person to the clarifications and keep the ban on sending. If he makes a mistake, he gets a second changed input and again looks for the basis of every field. If he copes on his own, the test of his written instructions with a colleague can be assigned.
Good speed does not prove the ability to check. Assigning the instructions test before he completes the second input independently would be premature.
What to give an experienced participant who already has a personal tool
For an experienced participant in a shared session, a useful task is to explain how the tool is built: where it gets its data, what it does with them, what it outputs, who can read the result and when the tool must stop. The task uses the same training material as the beginner's work, so the group works on one request at different depths. If a personal tool is built as an Agent Skill, that is, a package of instructions and files for a model, it is described using the same scheme. This is the author's recommendation for a session and is not a validated method.
Training material
Data source, read-only:
Request
Goal: prepare a review of the available training materials
Responsible person: session organizer
Date: 20.10.2026
Processing rules:
- The request may be read but not changed.
- The result is a draft of three fields: goal, responsible person, date. It is saved only to the folder of permitted outputs.
- If a field's value is not in the request, it is not invented. In the draft the field stays visible with the note "unknown, clarification needed", and the question is written to the list of clarifications. The action that needs this value waits for the organizer's answer.
- If there is no permission to write to the folder, writing stops without a workaround: no other folder is used, and the reason is reported to the organizer or administrator.
- Before any actual sending, the process owner's confirmation is needed. The owner's identity is unknown; this is a clarification for the organizer.
- Sending is prohibited in the training environment, and the exercise ends with a draft.
Task for the beginner
Prepare a draft from the request and check it with four questions. Do the three fields match the source, or are they honestly marked as unknown? Is the status "draft, not sent" given? Is the owner's confirmation written as a condition before a future sending, with no claim that it was received? Does the text contain any statement that the request was sent? The beginner does not describe the processing scheme; it is enough to name the place where the saved file is.
Task for the experienced participant
- Describe the scheme in a table: source, processing, result, who can read it, who can change it, and what permission each step needs. If the recipient of the result is only assumed, write it as such and note that their access still needs to be confirmed.
- Mark with the word "unknown" everything that the material does not state. No guessing.
- Walk through two failure cases. First remove the date from a copy of the request and check that the field stays in the draft with the note. Then walk through the missing permission to write to the folder. This may be reproduced only in a training environment that the organizer has separately approved for such a check. Permissions in a working system are not changed for the sake of the exercise. If there is no safe environment, limit the work to a written description of the expected behavior and mark the item "not checked".
- Write instructions for a colleague: what to open, what to enter, where the result is, when to stop, whom to ask. The colleague will follow them on a changed input without verbal hints.
Expected review
| Element | Expected description |
|---|---|
| Data source | A request of three fields. The tool and the participant read it, and it must not be changed |
| Processing | Checking the fields against the rules, preparing the draft, moving unknown values to the clarifications with a note in the field itself |
| Result | A draft in the folder of permitted outputs. The participant changes it. The assumed recipient is the process owner, but their right to read the folder is not confirmed and goes into the clarifications |
| Permissions | Reading the request and writing to one folder. No right to send. The material does not grant the process owner read access, so it must be confirmed with the organizer |
| Unknown | Who the process owner is, whether their access to the folder is confirmed, everyone with read access to the folder, the retention period for drafts |
In the failure case with no date, the draft keeps all three fields: the goal and the responsible person are taken from the request, the date field says "unknown, clarification needed", the question about the date is added to the list of clarifications, the value is not invented, and the status stays "draft, not sent". Everything that depends on the date, such as a deadline in the review text, waits for the organizer's answer. In the failure case with no write permission, writing stops, the reason is reported to the organizer or administrator, the file is not saved anywhere else, and the draft is not presented as a saved result.
What was done independently and what with help
The trainer records separately what the participant described alone, what was corrected after a specific hint and at which step, what was not checked, and what was not observed because of the environment. An example of such a note: "scheme and list of unknowns - independently; wording of the write permission - after the trainer's question; second failure case - not checked, no safe training environment." Written instructions are confirmed as good only when a colleague follows them successfully on new input.
Limits of the shared session
In the shared session, a participant may, if they wish, fill in a table of the same kind (source, processing, result, readers, permissions) for their own tool, but only in anonymized form: no confidential data, no code and no details that cannot be shown to the group. A detailed audit of a personal tool's code, a review of its architecture and individual integrations stay outside the shared session. They are done separately, individually, with the scope of work agreed in advance.
What the trainer does when the queue of help requests grows
When the trainer cannot keep up with checking outputs against the reference answer, the queue grows. Below are three typical situations, with lines composed for illustration.
"I can't save the answer." The participant writes in the chat: "I tried the floppy-disk button, clicked in the field on the right, and the answer did not appear in the permitted folder. I can log in to my account and the environment is open." The trainer asks what exactly the participant clicks and shows one step. If the problem repeats after the demonstration and the trainer becomes sure that the environment does not allow the result to be saved, the question goes to the administrator.
"I can't log in." The participant cannot access the environment with their account. Access is fixed first. Only then is a new observed attempt assigned on the same material. The technical failure is recorded separately from the skill.
"I don't understand what 'the process owner's confirmation' means." The trainer points to the condition stated in the source: the process owner's approval is required before sending. The source does not say who the process owner is, so the trainer asks the organizer and passes the answer on. Sending stays prohibited, and the participant still completes the task on their own.
A helper assigned to the session answers questions about specific stuck steps and leaves the task itself to the participant. The helper's access rights are not extended.
When the queue does not clear, part of the participants get an independent written block: they read the source, write down the conditions and compose a list of clarifications without a model. This is work on understanding the task, and no tool is needed for it. The other part moves to a subgroup agreed with the organizer, with a helper the organizer has approved.
In a group of up to 15 people, the need for a helper depends on the help load, the number of simultaneous obstacles, the available time and how well the instructions are prepared. There is no universal scientific limit here.
An area that could not be observed is marked "not observed" by the trainer. A participant who did not complete the practice task because the trainer was overloaded gets new time and new input.
Difficult situations: first action, check and next step
| Situation | First action | Check | Next step |
|---|---|---|---|
| No access to the environment | Have the administrator remove the block | A new attempt on the same material once access exists | A complete practice cycle |
| Does not understand the task | Walk through the source: fact, condition, exception | The participant names the conditions before opening the model | Working with the tool |
| Understands the task but cannot use the tool | Demonstrate one interface step | An independently saved output | Changed input |
| Runs the tool but does not check | Compare with the reference answer and look for the basis of each field | On a new input, the participant independently checks the output and rejects any invented fragment | Test of written instructions with a colleague |
| Can do everything, but the colleague cannot follow the instructions | Test the written instructions on the second input, with no verbal hints | The colleague completes the changed input | Revise the instructions, then repeat the test independently |
Several gaps at once are handled in turn: first the technical obstacle, then understanding the task, then the tool, then checking, then the written instructions for a colleague. A short complete practice cycle with the trainer looks like this: help is recorded separately for each area, and the new input is done independently.

What goes into a participant's individual learning record
The individual learning record is a document the trainer keeps for each participant. It records observations of specific actions. The document contains five groups of fields:
- the original task, the version of the input, the participant's goals, their role and access rights
- blocking technical obstacles, recorded separately from skills
- observations for each of the four areas (independently, after help, not completed, not observed) and an exact description of the help: what the trainer said, at which step, and what explanation the participant received personally
- evidence: a result file stored with restricted access, an independent attempt, and the outcome of testing the written instructions with a colleague
- the next task, the review date and the trainer's name
The help received is recorded so that the next attempt can be independent. Individual results are not shared with the group. Who has read access to the records and how long the training diagnostics are kept are agreed separately, before the cycle begins.
A manager in the group gets a separate learning record: defining the task and accepting the result. This is a different role with its own goals.
A home attempt counts as an observation when the organizer has approved the material and permitted the environment, the participant reports the help received, and the trainer reviews how the attempt went. When needed, the trainer asks for an observed repeat on the next input.
Frequently asked questions
Can a confidence questionnaire replace the practice task?
No. A questionnaire records a self-report, while a practice task records an action. A practice task on training material gives a concrete result that can be compared with the reference answer: the date is invented or moved to the clarifications, the responsible person is assigned or left in the clarifications. A questionnaire does not show this. The trainer uses the practice result to choose the participant's next task, and the questionnaire supplements the conversation with the participant without replacing observation.
What if a participant could not complete the practice task because they had no access to the environment?
The block is recorded as "not observed because of the environment", and no skill rating is given for that practice task. The access problem is fixed technically first. Once access is restored, a new observed attempt is held on the same material.
Can an experienced participant be allowed to help beginners?
Only with a clearly assigned role, without extending their access rights and without responsibility for other people's results. A good personal result relates to working with the tool, and handing over a process is a separate area of skill. The written instructions are tested by having a colleague follow them on changed input without the author's verbal hints.
How often should a participant's learning record be reviewed?
After the result of an independent attempt and after a discussion with the participant. If the independent attempt meets the reference criteria, the trainer assigns a more advanced task or a test of the written instructions with a colleague. If not, the trainer assigns the next specific help and new input. There is no fixed schedule, and the review is tied to the result of the attempt.
Does this approach suit a group of two and a group of fifteen?
The approach describes observations and next tasks, and it does not depend directly on the number of participants. The upper limit for a group in these sessions is 15 people. For a group of two and for a group of fifteen, the organizer decides in advance whether a helper is needed and whether sequential subgroups are needed. The decision depends on the load, the obstacles, the time and how well the instructions are prepared. If the queue of help requests grows, the organizer arranges a helper or splits the group into subgroups.
Can diagnostic results be used in performance appraisal?
No. The participant's learning record is a training document with a "not observed" category and a record of the help received. It records what the trainer saw in the session. It cannot be used in HR procedures unless the organization separately approves that use and changes the document's purpose in advance, in agreement with the participants.
How to start AI training in your own mixed group
You can start with one thing: the trainer sees the four areas of AI skill separately and assigns next tasks based on observed actions. A practice task on neutral material, changed input and an exact record of the help received are the three tools to begin the first session with. The formats of group sessions (up to 15 people, Russian and English, online or in person by arrangement) are described on the team training page.
Questions about the material: Telegram or hello@majento.ai.