7.1 Diagnose Underperforming Prompts and Outputs
7.1.1 Classify Before You Fix
Claude hands back a weak answer. The instinctive move is to hit regenerate and hope the next roll is better. Resist that instinct. A weak output is not a mystery to be re-rolled away — it is evidence, and the productive first move is to classify what kind of problem it actually shows before touching anything. Troubleshooting draws on everything else in this course: a bad output is almost always a prompting problem, a context problem, a model-fit problem, or a verification problem wearing a disguise, not an unexplainable failure.
That classification step matters because the fixes for each cause look nothing alike. Adding more detail to a prompt does nothing for a hallucinated citation. Switching to a bigger model does nothing for an under-specified ask. Distinguishing a knowledge gap, a prompt gap, and a context problem up front is what lets you respond with the fix that actually closes the gap, instead of guessing and burning another round trip.
Regenerating without changing anything just reproduces the same class of weak output; naming the symptom first is what makes the fix actually work.
The one idea to hold onto
Classify the specific symptom a weak output shows before choosing a fix. Troubleshooting is diagnosis, not repetition — the same under-specified prompt or the same crowded context tends to produce the same class of weak output on every retry.
7.1.2 The Symptom-to-Fix Table
Six recurring symptoms cover most weak Claude output, and each pairs with a specific, targeted fix rather than a generic "try again." Treat this table as a mental checklist to run through before you touch the prompt.
| Symptom | Likely cause | Targeted fix |
|---|---|---|
| Generic, vague, off-target | Under-specified prompt | Add audience, format, constraints, and an example |
| Wrong format / structure | Format not stated | Specify the exact output shape; provide a filled example |
| Confidently wrong facts / fake citations | Hallucination | Ground in source, require verifiable citations, allow "I don't know" |
| Quality drops late in a long chat | Crowded context window | Summarize, restart, or persist key info in a Project |
| Overkill cost/latency or too shallow | Model mismatch | Right-size the model — Haiku/Sonnet/Opus to the task |
| Missed part of a complex ask | No decomposition | Break the task into ordered, checkable steps |
Six symptoms, six distinct causes, six targeted fixes — the recurring diagnostic checklist for Task Statement 7.1.
Exam angle
Model mismatch is only one row in this table. A vague prompt, an unstated format, a hallucination, and a crowded context window are all fixed by something other than switching to a bigger, pricier model — upgrading the model leaves those four causes completely untouched.
7.1.3 Hallucination: Confidently Wrong, Not Just Wrong
When Claude states a wrong fact or invents a citation with total apparent confidence, that is hallucination — not a tone problem, not a phrasing problem. The targeted fix is to ground the answer in source material, require verifiable citations, and explicitly permit the model to say "I don't know" instead of fabricating an answer to satisfy the request. Allowing that admission is part of the fix, not a separate concession bolted on afterward.
7.1.3 — Exam Trap
Fixes that never touch grounding — raising temperature, asking Claude to "be more confident," shortening the prompt — leave the actual cause of a fabricated citation completely untouched. Only grounding in source and requiring verifiable citations targets hallucination directly.
7.1.4 Context Crowding: When Quality Drops Late in a Chat
A distinct symptom: output quality holds up fine early in a conversation, then degrades only near the end. That pattern points to a crowded context window, not a weaker model or a worse prompt. As a chat accumulates turns, low-signal content crowds out what actually matters, and the model effectively loses the thread it started with. The fix is to manage the context directly — summarize the conversation, start a fresh chat, or persist the key instructions and facts in a Project so they don't have to survive inside an ever-growing transcript.
7.1.4 — Exam Trap
Late-conversation degradation is easy to misdiagnose as "the model got worse" or "switch platforms." Adding more prohibitions to the prompt or downsizing the model does not address a crowded context window — summarizing, restarting, or persisting to a Project does.
7.1.5 Put It Together: Exam Traps for Task Statement 7.1
- •Regenerating repeatedly without changing anything. Random re-rolls are not troubleshooting — a repeated prompt tends to reproduce the same class of weak output.
- •Reaching for a bigger, more expensive model when the real problem is a vague prompt, an unstated format, a missing citation requirement, or a crowded context window. Model mismatch is only one row of the six — it does not fix the other five.
Where this shows up on the exam
7.1 questions describe a weak output and ask what to do next. Name the symptom silently before reading the options — generic/off-target, wrong format, fabricated facts, late-chat drift, overkill cost, or a missed sub-task each has exactly one matching fix in the table.
Key Takeaways
- ✓Classify the specific symptom of weak output before choosing a fix — do not just retry.
- ✓Generic/off-target output means under-specification; the fix is audience, format, constraints, and an example.
- ✓Confidently wrong facts or fake citations are hallucination; the fix is grounding and verifiable citations, not a shorter prompt or a 'more confident' instruction.
- ✓Late-conversation quality drops point to a crowded context window, fixed by summarizing, restarting, or persisting context in a Project.
- ✓Switching to a bigger/pricier model is a common wrong first move when the actual cause is a prompt or context problem.
- ✓Allowing Claude to say 'I don't know' is part of the hallucination fix, not a separate concession.
Check Your Understanding
Test what you learned in this lesson.
Q1.A user's prompt keeps returning generic, off-target drafts no matter how many times it's regenerated. What is the best first fix?
Q2.Claude confidently states a statistic and attributes it to a source that does not exist. Which fix targets the actual problem?
Q3.A long research conversation with Claude produces excellent answers for the first hour, then noticeably weaker, less relevant answers over the last twenty minutes. What best explains this and what should be done?
Q4.A workflow is producing shallow, surface-level output despite reasonable cost and latency, and the team's instinct is to upgrade to the largest available model. Before doing that, what should be checked first?
Practice This Lesson