Lesson 5 · 17 min
"Make the page good" is not a goal: the agent decides on its own that it is done. A verifiable goal is a command with the expected output, plus a check that stops it from being met by a shortcut. The agent runs in a loop until it passes. After that, what no command can prove (first-hand data, voice, usefulness) is left for the human gate, before the push.
Start the lesson
Goal: write the agent's goal as command → expected output, with a check against shortcuts, and only push after the human gate.
Tap each block.
The command proves what is mechanical; the gate proves what is valuable.
Tap each step.
Can you check it with a command? Is the answer yes or no? Is it impossible to meet without doing the work? Does the proof appear in the output? Are function, regression and limit covered?
These are the questions from the goal.md template in the Execução Longa (Long Run kit) (open, from INEMA).
Function: the page does what it should (one H1, right canonical, JSON-LD present). Regression: the rest of the site didn't break (tests and build). Limit: the agent didn't touch where it shouldn't.
On the INEMA portal, the rule is to run the tests and the build before the push: main has already broken because of a translation.
Of the 16-item checklist, part is mechanical (one H1, title of 50 to 60 characters, links that return 200, valid JSON-LD) and becomes a command.
Another part is judgment: verifiable first-hand data, human reading, swap test. That goes to gate 2, 10 to 20 minutes per page.
Codex and Claude Code have the /goal command: the agent works until the condition shows as met in the output. In Claude Code, the evaluator only reads the conversation, so the proof has to be printed.
Stop and step in if there are 3 cycles with no progress or if the agent reopens an item that is already done.
Swap the paths for the ones in your project. Anything between less-than and greater-than signs is for you to fill in.
# Goal — new page: <slug> ## Result The page /<slug>/ ready for gate 2, without breaking the rest of the site. ## Done criteria (command → expected output) **Function** - [ ] grep -c '<h1' dist/<slug>/index.html → 1 - [ ] grep -o 'rel="canonical" href="[^"]*"' dist/<slug>/index.html → https://www.seusite.com/<slug>/ - [ ] grep -c 'application/ld+json' dist/<slug>/index.html → 1 or more **Regression** - [ ] npm test && npm run build → no errors **Limit** - [ ] git diff --stat -- tests/ docs/CHECKLIST.md → empty - [ ] git status --porcelain → only the new page and the 2–3 pages of the link ladder ## Human gate (the command does not prove it) - first-hand data with a link to the evidence (checklist 2) - human reading: opinion, what went wrong, the author's voice (checklist 9) - swap test (checklist 10) ## Constraints - through the subscription only; no API without authorization - no push: the push is mine, after the gate
/goal The terminal output shows each command from goal.md with the expected output, and goal.md has all criteria checked off. Do not push. When you finish, stop and show me the diff.
Use a new page (or a rewrite) from your site.
Write the page's goal.md with at least one function criterion, one regression criterion, and one limit criterion, and run the five questions on each one.
Done when: each criterion has a command and the expected output, there is at least one check (a frozen file or a count with content) and the human gate section lists what the command does not prove.
Run the agent with the loop request. When it stops, read the page as a reader, decide, and only then do the push yourself.
Done when: the conversation shows each command with the expected output, goal.md has your gate 2 approval with a date, and the push went out after it.
The explanation appears after your answer.
Question 1 of 4
"The build passes with 0 failures" is a goal of which level?skill
Question 2 of 4
Which of these checklist items does NOT become a command and is left for the human gate?
Question 3 of 4
Why does the proof have to appear printed in the output when you use /goal in Claude Code?
Question 4 of 4
The limit criterion "git diff --stat -- tests/ → empty" protects against what?
Choose an option. The explanation appears after your attempt.