Increase autonomy based on evidence
What it is
The bicycle method represents supervision that changes as you learn. In the first runs, monitor everything. After you observe stability, review samples and maintain controls over actions with higher impact.
Why learn
Trusting more doesn’t mean removing all protection. A skill that produces a draft can operate with little supervision; publishing, sending, or editing records involves different consequences. The autonomy level depends on risk and the evidence.
Key concepts
Phase 1: track three different cases.
Phase 2: review each final delivery.
Phase 3: sample low-impact outputs.
Always: review blockers and behavior changes.
✓ Do it like this
Increase supervision when input or the environment changes.
✗ Avoid this mistake
Treat some good runs as unlimited authorization.
Practice before revealing
A stable skill starts using a new publishing tool. What changes?
View commented response
Capacity and risk changed. Reassess permissions, test cases, and supervision. The draft’s stability doesn’t prove publication safety.
Give feedback that identifies the smallest correction
What it is
Effective feedback names the observed result, the expected result, the impact, and the proposed change. Also include what worked and should be preserved. This reduces the risk that rewriting unnecessarily breaks good parts.
Why learn
“It’s bad” pushes the agent to guess. “The total is correct, but there’s a missing period in the header” pinpoints the defect. The correction is to add and test the period, not to replace the entire calculation pipeline.
Key concepts
Worked: totals by channel are correct.
Failed: the period doesn’t appear before the summary.
Impact: the reader doesn’t know which week it refers to.
Adjustment: require the period in the header and verify both dates.
From concept to action
- Observed: identify the initial condition.
- Expected: apply the described decision.
- Minimal correction: check the effect in the example.
- Test: record the output evidence.
✓ Do it like this
Cite the file, section, or a concrete example.
✗ Avoid this mistake
Ask “make everything better again” when there’s a localized defect.
Practice before revealing
Turn “the text doesn’t sound like me” into usable feedback.
View commented response
“You used long sentences and three superlatives that don’t appear in my examples. Preserve the facts, shorten those sentences, and use the two approved texts as references for tone.”
Turn the failure into a regression test
What it is
After fixing, save a case that would have detected the failure before. This test goes into the next evaluation. Also run the previous cases to ensure the new protection didn’t make valid inputs impossible.
Why learn
Fixing an execution is local; the test preserves the learning. For example, if a privacy rule started hiding all numbers, the regression should reveal that it also removed essential data that was previously public.
Key concepts
Failure: report without a period.
Case: CSV with different dates.
Test: header includes the earliest and latest date.
Regression: total and channels remain correct.
✓ Do it like this
See the test fail before trusting it protects the scenario.
✗ Avoid this mistake
Accumulate rules that are never exercised.
Practice before revealing
Empty input now gives a clear error, but the valid case stopped working. Did the fix finish?
View commented response
No. You fixed one case and introduced a regression. Adjust the condition and rerun both tests.
Version the recipe and preserve the reversal
What it is
Save skill changes in version control or in identified copies, with a short history of the reason. Record the version used in each experiment to distinguish improvement from instruction changes due to environment changes.
Why learn
A useful history lets you go back to the previous version when a fix worsens the result. It’s not enough to save “final,” “final2,” and “now-works.” Identify what changed and the evidence that justified the decision.
Key concepts
Project version: 1.0.0
Correction patch: 1.0.1
New feature: 1.1.1
Convention for this project: minor preserves the patch.
Always record reason and tests in the history.
✓ Do it like this
Use the versioning convention documented by the project.
✗ Avoid this mistake
Confusing a local convention with the universal semver rule.
Practice before revealing
What information should accompany “I adjusted the description”?
View commented response
The trigger that failed, the changed text, and the positive and negative requests that were reassessed. That way someone can understand why the change exists.
Conclude the project: weekly report
What it is
Assemble the skill, the synthetic CSV, the generator, negative cases, and execution evidence. The kit provides a functional starting point; your job is to run it, explain the result, and adapt a small rule with a corresponding test.
Why learn
The final project proves the full cycle: contract, implementation, verification, and improvement. Reading the SKILL.md is only one part. You need to open the artifacts, check the account, and observe how the skill reacts to an invalid input.
Key concepts
python3 scripts/gerar_relatorio.py dados/vendas.csv \
--outdir saidas/rodada-01
python3 scripts/testar_relatorio.py
# Then ask Codex to run it via the skill
# and compare it with the script’s result.
From concept to action
- Install the example: identify the initial condition.
- Generate report: apply the described decision.
- Test failures: check the effect in the example.
- Deliver evidence: record the output evidence.
✓ Do it like this
Compare direct execution with skill-guided execution.
✗ Avoid this mistake
Claim the Codex trigger was tested just because the Python passed.
Practice before revealing
Add a per-line value limit, document it, and create a case that exceeds it.
View commented response
An acceptable solution defines the limit, rejects the exceeding value with a clear message, and keeps the R$ 500 case approved. This limit is a choice of the exercise, not a universal commercial rule.
Review your skills portfolio
What it is
A healthy library contains tasks with clear boundaries, examples, and maintainable upkeep. Review similar skills, competing triggers, and obsolete rules. Consolidate duplicates when there’s evidence they represent the same process.
Why learn
The final skill isn’t writing more instructions: it’s maintaining a system where each procedure can be found, executed, and evaluated. Redo the review when the work changes, not only when the text seems outdated.
Key concepts
Final review
[ ] Approved output and clear contract
[ ] A piece of work and tested triggers
[ ] Freedom calibrated per step
[ ] QA with evidence
[ ] Configuration assessed
[ ] Feedback became a test
✓ Do it like this
Prioritize the skill that causes the most observable rework.
✗ Avoid this mistake
Increase instructions for each incident without removing contradictions.
Practice before revealing
Explain your skill to another person in 60 seconds: input, output, boundary, and evidence.
View commented response
If the person can locate the folder, run the example, recognize a failure, and review the QA, the package is teachable. The next step is to test it with new, properly prepared real inputs.
Check your understanding
What change closes the learning cycle?
What you take from this module
Deliver a skill, test cases, QA, and improvement history.
- Increase autonomy according to the evidence.
- Give feedback that identifies the smallest correction.
- Turn the failure into a regression test.
- Version the recipe and preserve the reversal.
- Complete the project: weekly report.
- Review your skill portfolio.
Next action: save the exercise in your learning lab and record what still needs review.