Bad output is not one thing and the repairs are not interchangeable. Name the class, then apply its repair.
Read the output in front of you against the middle column. The row you land on tells you what to change. Rerolling is not on the list, and neither is giving up.
| Failure | Looks like | Repair | Mechanism |
|---|---|---|---|
| Missing information | Confidently answers a different situation than yours | Supply what it couldn’t have | Lesson 2 |
| Under-specified | Valid, competent, not what you wanted | Add constraints or an example | Lesson 3 |
| Fabrication | Specifics that are wrong — numbers, names, citations | Ground it, or scope it to a source | Lesson 1 |
| Wrong shape | Right content, wrong length, format or register | Specify format, or show one | Lesson 3 |
| Lost the thread | Contradicts itself, drops earlier instructions | Fresh chat, restated brief | Lesson 2 |
| Insufficiently decomposed | Does all of it, shallowly | Break into steps; give it room to work | Lesson 1 |
Under-specified and wrong shape both come back competent and unwanted, which is why they blur. The test is not what you’d change — it’s what you’d keep.
Would you keep the substance and repackage it? Wrong shape — a container problem, and the content is already right. Would you throw the substance out and start again? Under-specified — a content problem, and no reformatting reaches it. The repairs overlap (one good example can fix either), which is why the classes keep blurring and why keep-or-discard is the question that stays sharp.
The reroll trap. Regenerating an unchanged prompt draws again from the same place. If the prompt is the problem, so is every reroll. Change something, or stop running it.
Ask it to diagnose itself — then treat the answer as a hypothesis. “What in my prompt led you to produce that?” is a useful move that almost nobody makes. What comes back is a hypothesis about your prompt, not a log of what happened. Good for generating leads. Not evidence. Test it before acting on it.
The seventh option: settings, not specification. A shallow answer to a genuinely hard task is sometimes fixed by letting Claude work harder — thinking on, more effort — rather than by rewriting anything. Check your specification first. Reaching for the dial before checking the spec is the exact habit this card exists to prevent, and it is much more often a spec problem.
If the answer turns on a fact that is private and appears in no source you can supply, nothing on this card reaches it. That is not a hard prompt, it is the wrong tool, and the correct move is to stop. A diagnostic that can never return stop is a motivational poster. You want calibrated persistence, not unlimited persistence.