Separate the recipe from who cooks
What it is
A skill is a reusable procedure that guides an agent through a task. In the video analogy, the result is the pancake, the agent is the cook, and the skill is the recipe. The file records decisions you don’t want to explain again with every request.
Why learn
Without a procedure, the agent has to fill in gaps: size, format, sources, and quality criteria. Recording these choices reduces avoidable variation. This doesn’t make generation infallible: different ingredients and unavailable tools still change the outcome.
Key concepts
Loose request: “Make my report.”
Contracted request: “Turn this CSV into a
weekly summary, with totals by channel, period, and pending items.”
✓ Do it like this
Name the input and show an approved output.
✗ Avoid this mistake
Promise identical results just for existing an SKILL.md.
Practice before revealing
Choose a task you repeated three times and describe the input and output in one sentence.
View commented response
Example: “From a weekly sales CSV, produce a Markdown report with totals per channel and questions for review.” A task with boundaries lets you test the skill.
Differentiate prompt, skill, rule, and tool
What it is
A prompt asks for something in this interaction. A skill stores a specialized procedure. AGENTS.md records durable project guidance. A tool carries out a concrete capability, like reading a file or querying a service. These pieces complement each other.
Why learn
When everything goes into the global rule, details of one task start influencing unrelated tasks. When everything stays in the prompt, you repeat instructions. Separating roles helps you find the correct place for each correction.
Key concepts
Prompt: “Make the September report.”
AGENTS.md: “Project documentation is in Portuguese.”
Skill: “Validate data, calculate, write, and verify.”
Tool: a CSV reader or a calculation script.
From concept to action
- Request: identify the initial condition.
- Guidelines: apply the described decision.
- Procedure: verify the effect in the example.
- Tools: record the evidence of output.
✓ Do it like this
Put the calculation rule next to the process that uses it.
✗ Avoid this mistake
Treat MCP as a synonym for skill or for automatic access.
Practice before revealing
Where should the rule “don’t invent missing numbers in the report” go?
View commented response
In the report skill contract, with behavior for missing data. If it’s a general project requirement, it can also appear in the project guidance without duplicating contradictory versions.
Find tasks that deserve a skill
What it is
The best first candidate is a frequent task with a definition of “good” that you can demonstrate. It doesn’t need to be complex. A video description, a text review, or a short report already lets you learn the whole cycle.
Why learn
Starting with “manage my business” mixes planning, writing, accounting, and publishing. You can’t locate the mistake. A smaller task produces an assessable artifact and gives faster feedback before connecting important systems.
Key concepts
Candidate A: review a paragraph → easy to compare.
Candidate B: prepare a local report → verifiable numbers.
Candidate C: operate the whole company → too broad a scope.
✓ Do it like this
Start with a local output that can be reviewed.
✗ Avoid this mistake
Link the first version directly to external sends or changes.
Practice before revealing
Compare two tasks using frequency, output clarity, and the cost of error.
View commented response
Prefer the one that combines repetition, clear output, and cheap review. A less frequent task can win if rework is high; justify the criterion instead of using a score without context.
Draw the input and output contract
What it is
The contract describes what needs to be delivered, what will be produced, and what happens when information is missing. In our lab, the CSV contains date, channel, and value. The output includes observed period, total, totals by channel, and pending items.
Why learn
Missing data should not turn into silent assumptions. If “value” doesn’t exist, the skill can’t calculate revenue. A useful response identifies the missing field, stops only the dependent step, and preserves the original material.
Key concepts
Input: UTF-8, date,canal,valor columns.
Value: decimal with dot; BRL currency defined by the exercise.
Output: relatorio.md + qa.json.
If there is no mandatory column: report the error and do not generate a report.
✓ Do it like this
Define the date format and currency unit.
✗ Avoid this mistake
Add values from different currencies without an explicit rule.
Practice before revealing
Add a stopping condition and a limitation to the contract.
View commented response
Stop if there is an invalid date; limit the analysis to descriptive totals, without inferring causes of sales changes. The contract should say what to do, not just list prohibitions.
Read the video’s promise critically
What it is
The six steps are the author’s practical proposal: start from the result, define the task, calibrate freedom, verify, compare configurations, and incorporate feedback. “Better than 99%” is the promotional title of the reference, not a measured result in this course.
Why learn
The video shows different outputs across models in a visual task. This is a useful case study, but it doesn’t establish a universal ranking. Observed times from that one run also don’t guarantee duration, price, or quality for your account.
Key concepts
Hypothesis: “A cheaper configuration is enough for summaries.”
Test: same cases, same rubric, more than one run.
Valid conclusion: “In this sample, it met the criteria.”
From concept to action
- Author’s account: identify the initial condition.
- Hypothesis: apply the described decision.
- Local test: check the effect in the example.
- Limited decision: record the output evidence.
✓ Do it like this
Identify which conclusions depend on the context.
✗ Avoid this mistake
Repeat quality percentages without an experiment.
Practice before revealing
Turn “the expensive model is always better” into a testable hypothesis.
View commented response
“In our image selection cases, configuration A has fewer images out of context than B, using the same evaluation criterion.” Now there is a bounded comparison.
Prepare your learning lab
What it is
You’ll work with synthetic data and local files. Download the kit from the homepage, open the folder in Codex, and keep input, output, and evidence separate. The exercise doesn’t require a paid API or a connected CRM account.
Why learn
Separating folders makes it possible to redo an execution without replacing the original. With each change, you can compare the previous report with the new one and verify whether the fix had any unintended effect.
Key concepts
laboratorio/
dados/vendas.csv
saidas/
evidencias/
.agents/skills/relatorio-semanal/SKILL.md
✓ Do it like this
Save a copy of the input used in each comparison.
✗ Avoid this mistake
Change data, model, and instruction at the same time.
Practice before revealing
Open the kit CSV and manually calculate the total of the four lines.
View commented response
The values 120,50 + 80,00 + 199,50 + 100,00 add up to R$ 500,00. This standalone calculation will be a reference to verify the first report.
Check your understanding
A skill needs to access a CRM. What provides that access?
What you take from this module
Choose a repeatable task and explain where the skill fits in.
- Separate the recipe from the cook.
- Differentiate prompt, skill, rule, and tool.
- Find tasks that deserve a skill.
- Draw the input and output contract.
- Read the video’s promise critically.
- Prepare your learning lab.
Next action: save the exercise in your learning lab and record what still needs review.