From prompt to role|INEMA.CLUB-PRO

Lesson 6 · 16 min

Train in a cycle: five, ten rounds and a list

Once the agent can do the work, the part almost nobody does comes next: training. Don't ask just once. Ask five, ten times, evaluate it the way you'd evaluate an employee, gather the corrections into a list and run it in a cycle until it fixes everything. It takes more time and more usage, but in the end you have a written standard procedure that works for any agent.

Start the lesson
Humanoid robot at a table with five identical sheets in a row; a person points to the sheet marked with a red tag.

In this lesson you will

  1. 1Why a single run fools you
  2. 2From evaluation to the list of corrections
  3. 3The written standard procedure
  4. 4The new-model test

Goal: run the same task in a cycle, gather the corrections into one list, and close out a written standard procedure.

01The training cycle

Tap or click each step. The cycle only ends when the list is empty.

Scroll the page: each step below lights up its part of the diagram.

  1. 1/6

    Run

    One good run can be luck; one bad run can be bad luck. Five to ten runs, with different inputs, show what the agent does every time and what it does only sometimes.

  2. 2/6

    Evaluate

    Look at all the results using the sheet's criteria and compare them with your example. Note for each run: "on task A I didn't like this, on B this was missing".

  3. 3/6

    List

    Put your notes together into a single list, with no repeats, from most serious to least serious. This is what you hand to the agent, not a pile of complaints.

  4. 4/6

    Fix

    Hand over the list, ask it to adjust, and run again. Cross off what passed, keep what failed. Repeat until the list is empty in every run.

  5. 5/6

    Standard

    With the list empty, ask the agent to write the standard procedure (a simple text): the steps, the rules that came out of the corrections, and the final test. It holds for the coming years and for any agent.

  6. 6/6

    The whole training

    Run several times, evaluate, list, fix in a cycle until it is empty, and write the procedure. The same mistake doesn't come back.

In my case (Nei, from INEMA), a good part of the day works like this: I delegate, I give feedback, and each correction becomes a written rule. In a way, I train the agent as much as it trains me.

02Five training habits

Tap or click each one.

1. In the chat

Each run in a new conversation of the same project: that way one run doesn't contaminate the other.

The list of corrections goes into the project instructions before the next run. A new conversation without the list in the instructions forgets everything.

In the terminalPut the list in AGENTS.md (or in a file it points to), and make each round a new session.

03Templates for the cycle

Copy and use. Replace the text between < > with your own.

Evaluation (document or spreadsheet)
# Evaluation · <task> · <date>
Criterion: the Measure from the sheet + the reference example

| round | input | passed the Measure? | what turned out good | what to fix |
|---|---|---|---|---|
| 1 | <case> | yes / no | | |
| 2 | | | | |
| 3 | | | | |
| 4 | | | | |
| 5 | | | | |
Request: run again after pasting the list into the instructions
The list of fixes is in the project instructions, from most serious to least serious.
Do this task following the list: <this round's input>
When you finish, tell me which items on the list you checked and how.
Request: write the standard procedure
The list reached zero in all 5 rounds. Write the standard procedure for this task (a plain text that I will keep) with:
- when to use it and what it delivers;
- the steps, in order;
- the rules that came from the fixes (one per line);
- the final test: how to check against the Measure before delivering.
Use language that another agent, or another person, can understand without asking me.
Simulation: illustrative example
  1. The start date appears in the first sentence. (failed in 3 of 5)
  2. No student names, not even in examples. (failed in 1 of 5)
  3. The text fits on one phone screen. (failed in 2 of 5)

After the cycle: 5 of 5 on every item.

04Practice: train your agent

Use the sheet from lesson 4 and the goal from lesson 5.

E1 · Five rounds

40 min

Run the task from your sheet five times, each in a new project conversation and with a different input, and fill in the evaluation (document or spreadsheet).

Done when: the table has five filled-in rows, with "yes" or "no" in the Measure column.

E2 · The list of fixes

15 min

Gather what to fix into a single list, one checkable fix per line.

Done when: the list has at least three fixes, and each one can be checked by looking at the result.

E3 · The cycle until zero

40 min

Paste the list into the project instructions, run the same inputs again (new chats), and repeat until they all pass.

Done when: in one final round, all five inputs passed every item on the list.

E4 · The procedure

15 min

Ask for the standard procedure and keep it with the sheet (in the chat, in the instructions or as a project file; in the terminal, in the folder).

Done when: the file exists and has the rules that came out of the corrections and the final test.

05This lesson in one minute

Train the agent like an employee.

Train the agent like an employee (video in Portuguese) · 53 sVideo page with script →

06Test what you learned

The explanation appears after your answer.

Question 1 of 3

Why run the same task five to ten times?skill

07The lesson in three sentences

Several roundsFive to ten runs with different inputs show what the agent does every time.
List and cycleVerifiable corrections, one per line, applied in a cycle until none are left.
Standard procedureThe result of the training is a text that works for any agent and any new model.

Go further

The reel on YouTube
Previous lessonNext lesson