Lesson 2 · 18 min
In the kit's pilot, with a level 3 criterion and 14 frozen tests, the run finished in the first cycle, in 3 minutes. Most of the work in a long run is writing the criterion. Here you write your own goal.md so that the only way to meet it is to do the work.
Start the lesson
Goal: write a goal.md with command → output criteria that cover function, regression and limit and that can't be met by a shortcut.
A good goal.md has these five parts. Tap each one.
Function + regression + limit, each with a lock, and the proof visible. The full 0–4 scale is in course 0; how to build the verifier, in course 1.
Each "no" lowers the criterion's level.
"Make it good" and "reviewed and clean" have no command. Rewrite until you have one.
python3 validar.py && echo OK → OK
"Improved performance" is a comparison. Fix the number: "p95 under 200 ms on benchmark X".
Look for the cheapest cheat: deleting a test, an empty page, a hardcoded value. For each one, a lock: minimum count, hash or git diff of the tests, content validator.
"0 failures" becomes "0 failures AND ≥ 48 tests AND tests/ untouched".
For Claude's /goal, write the condition as something the agent prints: "the npm test output shows 0 failures". "The state.md says done" only counts if the agent shows the file.
All three, each with at least one line. And two tests: a quick one per cycle (seconds or minutes) and the full one at the end.
Filled-in example for a small task (counting words). Swap in your own.
# Goal — contador-palavras - **Início:** 2026-10-19 09:00 · **Agente:** Claude /goal - **Tetos:** tempo 2 h · memória 2G ## Resultado `conta.py` tem conta_palavras(texto) que conta palavras separadas por qualquer espaço. ## Critérios de pronto (verificáveis) **Função** - [ ] `python3 -m pytest -q` → `3 passed`, 0 failed (e nunca menos de 3 testes) **Regressão** - [ ] `python3 -c "import conta"` → sem erro (o módulo continua importável) **Limite** - [ ] `git diff --exit-code -- tests/ pytest.ini` → código 0 (testes congelados) - [ ] `git diff --stat` → só `conta.py` e a pasta longrun/ **Teste rápido por ciclo**: `python3 -m pytest -q` **Teste completo no final**: `python3 -m pytest -q` ## Restrições - só pela assinatura; nenhuma API - não mexer em: tests/, pytest.ini ## Portões humanos (parar e perguntar) - qualquer mudança fora de conta.py e da pasta longrun/
/goal The terminal output shows <comando de verificação> passing with <resultado esperado>, and longrun/<pasta>/state.md says "concluído".
/goal RESULT: <final result in one sentence> VERIFICATION (it only ends when all of these pass): - <comando> → <saída esperada> CONSTRAINTS: <what is off limits>. Subscription use only; no API without authorization. STATE: use longrun/<pasta>/ (goal.md, plan.md, state.md, progress.md, failures.md, decisions.md, canal.md). After any compaction or resume, reread goal.md, state.md, plan.md and canal.md before acting. Record in canal.md (append only) facts, lessons, glossary and traps as soon as they appear. CYCLE: analyze → pick the next useful action → execute → test → observe → fix → update state/progress → commit at the checkpoint → continue. As long as there is a next action that is objective, safe and aligned with the goal, execute it without waiting for new instructions. 3 cycles without measurable progress = stop, record it in failures.md and call me. STOP AND CALL ME ONLY IF: credit/API spending, irreversible or external action (push to production, email, delete), missing credential, business decision, real conflict between requirements.
From the kit's longrun/2026-10-01-medir-sessao/goal.md, the run that built medir-sessao.py:
Função - pytest -q → todos passam (12 rápidos + 2 com sessões reais), 0 falhas, 0 skips Regressão - python3 tools/medicao/cx.py | head -1 e cc.py continuam rodando Limite - git diff --exit-code -- tests/ pytest.ini → vazio (testes congelados; hash inicial 3ae1ba7975508cba) - só criar/alterar tools/medir-sessao.py e a pasta longrun/ Teste rápido por ciclo: pytest -q -m "not real" Teste completo no final: pytest -q
Result: done in the 1st cycle (3 min). Independent check afterward: test hash intact, no hardcoded values in the code, 14 passed reproduced, and a never-before-seen session checked.
List from the kit's templates/goal.md.
Use the folder you created in lesson 1.
Fill in goal.md with the result, function, regression, limit, the two tests, constraints and human gates. Before any work, run each criterion in the terminal.
Done when: every criterion has a command that runs now; the function ones fail before the work (if they already pass, the criterion doesn't measure the task) and the limit one already passes: git diff --exit-code -- tests/; echo $? → 0.
On a throwaway branch (git switch -c ataque), pull the cheapest cheat: delete a failing test, or swap the function for a fixed value. Run the criteria again and go back with git switch -.
Done when: for each cheat, at least one criterion fails; when you delete a test, git diff --exit-code -- tests/; echo $? → 1.
Question 1 of 4
The criterion is "pytest -q → 0 falhas". Which cheat does it let through?skill
Question 2 of 4
In Claude Code's /goal, why does the condition have to show up in the output?skill
Question 3 of 4
Before any work, you ran the function criterion and it already passed. What does that tell you?
Question 4 of 4
Which layer answers "didn't touch what it shouldn't and didn't spend what it shouldn't"?
Pick an answer. The explanation shows up after you try.