Examples, tests or interview?
Claude's work keeps coming back not quite right, and every answer option is a real way to iterate: better words, examples, tests, questions, one message or several. The trap is picking a good technique for the wrong gap. Name what's missing first, and the technique follows.
D3 Claude Code ConfigurationExam guide 3.5About 15 minutes
Start here: the hairdresser and the tailor
Three everyday ways of getting what you want from someone skilled.
Claude works the same way. Input/output examples are the photo: a few “this goes in, this comes out” pairs. Tests first is the chalk: you write tests for the answers you already know, Claude runs them, and you share the failures. The interview pattern is the tailor asking: Claude asks you questions before it writes anything. And when you have several problems, ask one more thing: do the fixes affect each other? If they interact, send them in one message; if they are independent, send them one at a time.
- Claude doesn't understand the transformation (a format, a conversion, a mapping): show concrete input/output examples.
- The right answers are known, but each fix breaks another case: write tests first, then share the failures.
- Nobody has said exactly what's wanted: the interview pattern. Claude asks before it builds.
- Several problems whose fixes affect each other: send them in one message. Independent problems: one at a time.
Words in the question that tell you which side you're on
Examples
- convert this format, transformation
- “the prose description keeps being misread”
- “rewritten the spec four times”
Tests first
- “each fix breaks another case”
- “the expected results are known”
- regressions, “checked by hand”
Interview
- “the request is vague”
- “requirements are unclear”
- “several reasonable readings”, “hadn't thought of”
One message or several
- “the problems interact”, “depends on”
- “independent issues”, “unrelated”
Try it: iterate until it's right
Pick a situation, then pick how you work with Claude. The simulator plays out each round: one square per case, green when right, pale pink when wrong, and dark pink with an outline when a case that was right has broken. The diagram above highlights the situation and your technique.
The rounds and minutes are illustrative, chosen to show the shape. They are not measurements. The exam guide's point is that each technique fixes one kind of gap: examples for understanding, tests for stability, an interview for unclear requirements, and one message or several depending on whether the problems interact.
Things to try: addresses + prose, then addresses + examples. Then fare rounding + one message per issue to watch the breaks, and fare rounding + tests first. Then reminders + tests first, a good habit in the wrong order. Last, checkout bugs + one message per issue against unrelated bugs + one message per issue: the same choice, opposite verdicts.
The four traps
| Trap | What the wrong option looks like | Why it loses |
|---|---|---|
| Only usually works | “Rewrite the instructions in more detail, describing the transformation step by step.” | It can work, but concrete input and output examples communicate a transformation better than more prose. |
| Delays the problem | “Keep reporting each failing case to Claude as you find it.” | Each fix can break an earlier case. Writing the tests first and sharing the failures keeps every case in view. |
| Only usually works | “Write the spec yourself and hand it over before any work starts.” | Your spec can still miss what you haven't thought of. The interview pattern lets Claude ask about the gaps. |
| Different problem | “Send three interacting bugs one at a time, in separate messages.” | Fixes to interacting problems collide. Interacting problems go in a single message; independent ones in sequence. |
Not to be mixed up with Plan or just do it?, which is about whether to plan before editing. This page is about how you iterate once the work is under way.
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 3.5: “Concrete input/output examples communicate transformations better than prose.” “Test-driven iteration (write tests first, share failures); the interview pattern (Claude asks questions first); single message for interacting problems, sequential for independent ones.”