Build & submit taskBetaintermediate

Diagnose Four Failing Claude Prompts Down to Root Cause

Take four prompts that genuinely underperform, and for each one work from symptom to hypothesis to a named root cause, then fix it and paste the retest that proves the fix held. One of the four must turn out to need something other than a prompt edit. Closes with an efficiency optimization measured in rounds and elapsed time. No coding required.

2 hrs

Est. time

5

Outcomes

8

Rubric criteria

65%

Pass score

What you'll learn

Skills you'll have real reps in after shipping this.

Name the symptom before reaching for a fix
Too generic, invented specifics, and constraint ignored are three different failures with three different repairs. Rewriting without classifying is how the same problem returns in a new form.
A rejected hypothesis is part of the evidence
Considering two explanations and saying why one lost is what separates a diagnosis from a guess that happened to work.
Sometimes the prompt is fine
Context that has grown too long, a missing knowledge source, a surface that does not fit the job, or a model tier mismatched to the difficulty all produce symptoms that look like prompt problems and do not respond to rewriting.
Retest or it did not happen
A fix without a pasted retest is a claim. Running the same failing case again is the cheapest possible verification.
Optimization has a unit
Rounds of back-and-forth and elapsed time make an efficiency gain something you can defend, where "it feels smoother" does not.

See how it works

When the prompt is not the problem

Lost in the middle
Where the fact sits in the prompt changes whether the model finds it.
FACTToken 1Token 64,0000%25%50%75%100%Position of the test fact within a 64k-token promptRecall accuracy
Where the fact sits
Beginning of the prompt
Model recalls the fact
94%
Recall accuracy by needle position on a 64k-token prompt. Numbers are representative of published needle-in-haystack benchmarks; the U-shape is robust across models.

Instructions buried in a long conversation lose their grip on the output, which looks exactly like a wording problem and does not respond to rewriting. Testing the same prompt in a fresh conversation separates the two in one step.

The scenario

When a prompt underperforms, the reflex is to rewrite it and try again. Sometimes that works, and when it does nobody learns why, so the same failure returns next week wearing different clothes. The prompts that resist rewriting are the ones where the prompt was never the problem: the conversation had accumulated so much material that the relevant instruction stopped carrying weight, the task needed knowledge that was never provided, the work belonged on a different surface, or the request was asking for something no amount of phrasing would produce.

Diagnosis means naming the failure before reaching for a fix. The output was too generic, or it invented specifics, or it ignored a constraint you stated, or it answered a different question than the one you asked, or it was right but unusable in the format it arrived in. Each of those points at a different cause and a different repair. This task makes you run four real diagnoses to root cause, verify each fix with a retest you paste, and include one case where the answer turned out to be a different surface, a different model tier, a fresh conversation, or knowledge the prompt never had.

Your role

You are the person colleagues bring their broken prompts to. Your deliverable is four diagnoses that go from symptom to root cause to a verified fix, plus one optimization that made a working process measurably cheaper to run.

Start the task to unlock the full brief

You'll get the step-by-step requirements, setup commands, the 8-criterion grading rubric, tips, and the ability to submit your solution for instant AI grading.

Free to start · submit when you're ready

What you'll build in this prompt troubleshooting task

This is a build-and-submit task rather than a guided lab, and it requires no coding. You define a short failure taxonomy, then diagnose four prompts that genuinely underperformed, working from the symptom through competing hypotheses to a named root cause. Each diagnosis ends with a fix and a verbatim retest, because a fix without a retest is a claim.

The requirement that makes this more than rewriting is that one of the four has to turn out not to be a wording problem at all. Context that grew too long, a knowledge source that was never provided, a surface that did not fit the job, or a model tier mismatched to the difficulty all produce symptoms that look like prompt failures and do not respond to rewriting. The document closes with an efficiency optimization on a process that already worked, measured in rounds of back-and-forth and elapsed time, and a specific first check to run next time.

Grading is rubric-based and explainable. Your submission is scored against weighted criteria covering the taxonomy, the verbatim evidence, the competing hypotheses, the verified fixes, the non-prompt root cause, and the measured optimization, with per-criterion feedback quoted from your document. The pass threshold is 65 percent and you can resubmit. Troubleshooting and optimization is a scored domain on the Claude Certified Associate Foundations exam.

Frequently asked questions

How is this different from the prompt iteration task?

The iteration task improves prompts that already work, changing one variable per round to find out what earns the gain. This one starts from failures and works backward to root cause, including cases where the prompt was fine and something else was wrong. Different slice of the same skill set.

Do I need coding or API access?

No. Everything runs in claude.ai, including the free tier, and the deliverable is a Markdown or PDF document. No terminal, API key, or development environment is involved.

Why require two hypotheses per failure?

One hypothesis that turns out to be right is a guess that landed, and it teaches you nothing about the next failure. Considering an alternative and saying why you rejected it is what makes the reasoning transferable, which is why the rubric weights it heavily.

What if I cannot find a failure that was not a wording problem?

Take a prompt that failed inside a long conversation and run it word for word in a fresh one. If the output improves, the context was the cause rather than the wording, and that single test is usually enough to surface the case the task is asking for.

What counts as a complete submission?

One Markdown, text, or PDF file with a named failure taxonomy, four real diagnoses each carrying the verbatim prompt and bad output, a taxonomy classification, two or more hypotheses with rejections explained, a root cause, a fix, and a pasted retest, at least one non-wording root cause, an optimization measured in rounds and elapsed time, and a specific first check for next time.