Lesson 7 · 18 min
The overnight queue is only worthwhile if the morning is short. Read the report in the chat, check each result against what you expected, note how many were right on the first try, and turn the task that failed most into a new rule.
Start lesson
Goal: review the overnight batch, measure first-try accuracy, and write a rule for the task that failed most.
Fifteen minutes to improve the next night.
Scroll down: each step below lights up its part of the diagram.
Each task replied in the chat when it finished. Start by reading the conversation from top to bottom; /jobs and /completos show what finished, and /status shows what is still open.
Open each result and compare it with the "Checkable result" column in your paper queue. The question is simple: can you use it as is?
Count how many were right without any changes from you. This number, recorded every morning, shows whether the queue is improving or just running.
Take the task that failed most and write a one-line rule that would have prevented the error. Add it to the project's AGENTS.md or the skill prompt. Run it again and compare.
Read, check, measure, and turn the error into a rule. The agent does the work; you improve the process.
Click each block. Fixing it by hand solves today; the new rule solves the next nights.
Read the overnight chat. For each task, you should have a result or a reason for failure. A task with no reply is still running or waiting: check /fila.
/completos lists the flows that finished; /status shows the ones that still need something from you (running or failed).
Open the file or text and compare it with what you expected. Mark it: right on the first try, right with adjustments, wrong, or failed.
Don't fix anything yet. First look at the whole batch: the pattern in the errors appears when you see them all together.
Read the reason. código 143 means it restarted, not that the agent made an error: /refazer. For an input error (broken link, wrong file), fix the input and send it again.
If the agent got it wrong, doing the same thing again isn’t enough: you need a new rule.
A good rule is short, concrete, and testable. It says what to do, not just what to avoid. Add it to the project’s AGENTS.md (step 1) or the skill prompt.
Run the same task again with the rule and compare. If it improved, keep the rule; if not, rewrite it.
Meeting summaries: start with decisions made and who is responsible, in no more than 8 lines.
One line per task. Keep each morning’s results in the same file so you can see how things change.
## Morning of YYYY-MM-DD | # | Task | Result | Right the first time? | What went wrong | Action | |---|---|---|---|---|---| | 1 | Transcribe Monday’s meeting | text | yes | — | — | | 2 | Queue explainer | video | no | too long | new rule | | 3 | | | | | | Right the first time: __ of __ Task that failed most: __ New rule: __
Read conferencia.md. For the task that failed most: 1. Describe the error in one sentence, with an example of the result. 2. Propose ONE concrete, testable, one-line rule for AGENTS.md or the skill prompt. 3. Tell me how to check tomorrow whether the rule worked. Don’t change any files; only make a proposal.
/completos /jobs /espaco /limpar artefatos 14 # only shows what would be removed (more than 14 days old)
Right the first time: 7 of 10. Two were right after an adjustment (summaries were too long), and one failed because of a broken link.
New rule: "Meeting summaries: decisions and who is responsible first, no more than 8 lines."
The next night: 9 of 10.
The morning after E4.
Fill in conferencia.md with every task from the night: the result, whether it was right the first time, and what went wrong. Address failures based on their cause.
Done when: each task in the batch has a line, and you noted how many were right the first time.
Write a rule for the task that failed most (in AGENTS.md or the skill prompt). Run the same task again the next night and compare.
Done when: the rule is written, the task ran again, and you noted whether the result improved.
Extra challenge for sequence 3: 03-filas.md.
Choose a start date and how much time you have each week. The plan accounts for exercises that require a night with the queue.
Where step 3 fits between the terminal, your Jarvis, and automated social media.
The explanation appears after your answer.
Question 1 of 3
Which number should you write down every morning?skill
Question 2 of 3
A summary came out too long for the third night in a row. What should you do?
Question 3 of 3
A task failed with código 143. What should you check?
Choose an option. The explanation appears after you try.