Choose an example that deserves to be repeated
What it is
Reverse engineering starts with an output you consider good and the inputs that produced it. An isolated report shows the appearance, but doesn’t explain filters, formulas, and decisions. Also save the data’s origin.
Why learn
Copying only the format can reproduce an elegant mistake. Before turning the example into a reference, check numbers, dates, and context. A good reference example has known quality, not just approval from general impressions.
Key concepts
Expected report
Observed period: 2026-09-14 to 2026-09-17
Total: R$ 500,00
Store: R$ 320,00 | Site: R$ 180,00
Pending items: none in the example file.
✓ Do it like this
Confirm the example’s sum before using it as the answer key.
✗ Avoid this mistake
Choose the prettiest output without knowing the data.
Practice before revealing
List three mandatory characteristics of this report.
View commented response
Explicit period; exact sum of R$ 500,00; totals by channel that reconcile with the overall total. The way you write the introduction can vary.
Interview the process, not just the style
What it is
Ask how the work was done: which sources were used, which records were excluded, how duplicates were handled, and who approved the result. Ask questions that make hidden decisions explicit, rather than asking only for a “professional style”.
Why learn
Someone who has done a task for years takes shortcuts without noticing them. The agent doesn’t know those shortcuts. A short interview recovers tacit knowledge and helps uncover dependencies that must exist before automation.
Key concepts
1. Does this total include cancellations?
2. Which date defines the week: purchase date or payment date?
3. Two identical records are duplicates or two sales?
4. What do you check before sharing?
From concept to action
- Source: identify the initial condition.
- Filter: apply the described decision.
- Calculation: verify the effect in the example.
- Presentation: record the output evidence.
✓ Do it like this
Record “we don’t know” when the answer doesn’t exist.
✗ Avoid this mistake
Deduce the business rule by looking at a single line.
Practice before revealing
Create a question that prevents improper refund summation.
View commented response
“Do negative values represent authorized refunds? Should they subtract from the total or appear in a separate section?” The answer becomes an explicit business rule.
Convert preferences into acceptance criteria
What it is
“I want something good” does not allow you to consistently reject an output. Break quality into observable items: the presence of the period, alignment between the data and the table, the summary length, and the clarity of the outstanding points. Separate mandatory requirements from preferences.
Why learn
Up-front criteria prevent changing the bar after seeing which model produced the result. They also help distinguish a serious block, like a made-up number, from an editorial improvement, like a long sentence.
Key concepts
Criteria | Evidence
Correct total | sum independent of the lines
Explicit period | header with two dates
Clarity | reader identifies total and pending items without searching
No invention | statements traceable to the input
✓ Do it like this
Define blocking failures before the first evaluation.
✗ Avoid this mistake
Approve based on the number of checklist items without looking at the content.
Practice before revealing
Rewrite “well organized report” as two criteria.
View commented response
“Total and period appear before the comments” and “open items stay in their own section, with a proposed action.” Both make the organization not inspectable.
Separate what is fixed from what can vary
What it is
The template must preserve the necessary information and allow adaptation where it helps. In the report, column names and sum rules are fixed; the wording of comments depends on the content. This separation is the foundation for the level of freedom.
Why learn
Freezing everything produces generic text. Leaving everything open allows important numbers to be omitted. A good recipe combines a stable skeleton with delimited editorial decisions—without trying to predict every sentence.
Key concepts
FIXED: period → total → channels → pending items.
VARIABLE: observations about concentration by channel.
PROHIBITED: “the campaign caused the increase” without campaign data.
✓ Do it like this
Provide examples of two different writings that would be acceptable.
✗ Avoid this mistake
Require the same conclusion when the data changes.
Practice before revealing
If there’s only one channel in the file, should the skill create a comparison between channels?
View commented response
No. It tells you there’s a single channel and limits the analysis. Preserving the contract means adapting the comment to the input, not fabricating variety.
Build your first reference cases
What it is
A reference set brings together inputs, expected outputs, or criteria, and the reason each case exists. Include the common path, a missing data point, and an ambiguous situation. The set represents the task, not just your favorite example.
Why learn
A correction that resolves the original file can break an empty file. Varied cases reveal this effect early. Reserve a few cases for final evaluation, without using them to edit the instruction on every attempt.
Key concepts
C01: four valid sales → total 500,00.
C02: no value column → clear error, no report.
C03: empty CSV → report that there is no data.
C04: negative value → follow the explicit policy.
From concept to action
- Normal: identify the initial condition.
- Absent: apply the described decision.
- Ambiguous: check the effect in the example.
- Reserved: record the output evidence.
✓ Do it like this
Note why each case is in the set.
✗ Avoid this mistake
Call a single definitive benchmark run.
Practice before revealing
Add a case that detects a decimal separator error.
View commented response
A line with the value “120,50” must be rejected by the contract using a decimal point, or normalized by a deliberate rule. In the kit, the decision is to reject with a clear message.
Write the briefing that turns into a skill
What it is
The briefing gathers purpose, inputs, output, steps, limits, and criteria. It doesn’t yet need to be an installed skill. First, check whether someone else can execute the process using only these instructions, without relying on you at every step.
Why learn
This exercise surfaces ambiguities before you blame the model. If even you can’t say when the work ends, automation just speeds up the spread of uncertainty. Use a small, complete briefing as a starting point.
Key concepts
Transform the provided CSV into a local weekly report.
Validate date, channel, and value; calculate without changing the input.
Produce period, total, channels, and pending items.
Check the sums and deliver relatorio.md + qa.json.
If a required field is missing, explain and stop the calculation.
✓ Do it like this
Ask the agent to list ambiguities before formalizing.
✗ Avoid this mistake
Ask for “create the perfect skill” without providing a reference.
Practice before revealing
Use the kit’s briefing template to describe your own process.
View commented response
The delivery must include an example input, a desired output, and at least three criteria. If a business rule is missing, record the open item instead of completing it by guessing.
Check your understanding
You received a nice report, but without the period. What’s the diagnosis?
What you take from this module
Produce an output specification and an initial set of cases.
- Choose an example worth repeating.
- Interview the process, not just the style.
- Turn preferences into acceptance criteria.
- Separate what is fixed from what can vary.
- Build your first reference cases.
- Write the briefing that turns into a skill.
Next action: save the exercise in your learning lab and record what still needs review.