Three sentences you will hear every day
“I have fixed the issue.”
“All tests pass.”
“This should work now.”
Read them carefully. Not one of them is evidence. They are claims.
The first two might be true. The third contains the word “should”, which is doing an
enormous amount of quiet work.
Why the agent talks like this
Back to Lesson 1.3. There is no small voice inside saying “I am not sure”. The confident
tone is a writing style learned from documentation and tutorials, which are also written
confidently.
An agent has one advantage over a chat window here: it can actually run things. But it does
not always run things, and when it does, it does not always read the output carefully.
Replace questions with requests for evidence
This is a small change in how you talk, and it changes everything.
| Instead of asking | Ask for |
|---|---|
| “Did that work?” | “Run the tests and show me the output.” |
| “Is it fixed?” | “Show me the failing case working now.” |
| “Are you sure?” | “What did you actually run, and what did it print?” |
| “Does it handle empty input?” | “Show me what happens with an empty list.” |
You are not being difficult. You are asking for the thing that would convince you, which is
what review is.
The habit that saves you
Before you accept “done”, ask yourself: what did I actually see?
If the answer is “I saw a message saying it was done”, you have seen nothing.
This sounds harsh. In practice it takes about four seconds, and it is the difference between
the two Module 4 bugs being caught in a minute or living in your app for a month.
One more thing to watch
Sometimes an agent will report that tests pass when some were skipped, or when it ran only
one file. The output is true; the summary is misleading.
Read the numbers. “12 passed” is useful. “Tests pass” is not.
Try this before the next lesson
- Find a place in your Recall session history where you accepted “done” without seeing
anything. Go back and check it now. - Ask your agent to prove that
nextGap(200, "easy")returns 365. Notice what it does. - Ask it to run only one test file and then say whether the project passes. Watch how
easily a true statement becomes a wrong impression.