Troubleshooting and Optimization: CCAO-F domain 7 study guide

CCAO-F · Troubleshooting and Optimization (10% of the exam)

Troubleshooting and Optimization is domain 7 of the Claude Certified Associate Foundations exam, 10% of the scored items. It tests whether you can find the real cause when Claude gives poor results, fix it with a change you can measure, and make a working Claude process faster or less wasteful without losing quality.

What the official guide covers

The Claude Certified Associate Foundations exam guide (version 1.0, effective July 2026) lists three tasks under domain 7, Troubleshooting and Optimization:

What the guide lists	What it means in practice
Identify, diagnose and resolve issues with underperforming prompts or poor outputs	Trace a bad result to its cause: the prompt, the context, the feature or model, the setup or the task itself
Adjust approach based on feedback and results	Change one thing, test it on the same inputs, keep what works
Optimise workflows for efficiency and effectiveness	Remove repeated uploads, wasted back-and-forth and over-powered settings

Diagnose in this order

Most poor outputs have one of five causes. Check them in order, because the early ones are the most common and the cheapest to fix.

The prompt. Missing reader, source, format or a rule for gaps.
The context. Wrong, old or too much material; a long chat that has drifted.
The feature and model. A calculation written out in prose rather than run with code execution, or a model too light for the reasoning needed.
The setup. Project instructions, knowledge files, Skills or connectors out of date or not doing what you think.
The task. Something Claude should not be doing alone, which needs a person or an escalation.

The order runs cheapest fix first. The common mistake is to jump straight to "use the most capable model" or "Claude cannot do this".

Read when the problem started

Similar-looking bad output can have different causes. When it started tells you which:

Pattern	Likely cause	Fix
Wrong from the first reply	The prompt left out context, limits or format	Add what was missing
Fine at first, worse as the chat grew	Context: earlier detail was summarised	Restart from a checked summary, or move the rule into project instructions
The same kind of error every time, such as figures slightly off or shallow analysis	Wrong feature or model	Code execution for numbers; a more capable model for depth
It worked last month	Stale setup: a file, instruction or Skill out of date	Run the Project maintenance review
The task asks for something no setup can give, such as next quarter's exact sales	Expectation mismatch	Reshape the task: ask for a range with stated assumptions

Symptom, cause, fix

Symptom	Likely cause	Fix
Generic output that could fit any company	No context about reader or purpose	Add the reader, purpose and an example
Answer is too brief	The prompt was framed as a quick question	Ask for the sections and depth you want
Figures that are not in your data	No source limit or gap rule	Limit to the source; tell Claude to mark gaps
A Project ignores a rule	Rule written in the name or description, which Claude cannot see, or conflicting instructions	Move it into project instructions; remove the conflict
Answers quote an old policy	Outdated file still in project knowledge	Remove the old file; keep one current version
Claude forgets a decision made earlier in a long chat	Earlier messages have been summarised to make room	Start a new chat from a checked summary
Claude says it "sent" or "saved" something that never arrived	No connector for that tool; the claim is a hallucination	Connect the tool, or do the action yourself
Claude does not use your Drive files	Connector not switched on for this chat or not signed in	Turn it on from the "+" menu and sign in
Shallow reasoning on a complex task	Model or effort too light	Use a more capable model, higher effort or thinking
Simple, repeated task is slow	Heavier model or effort than the task needs	Use a faster model at default settings

Worked example: fixing a prompt one change at a time

An events coordinator's prompt drafts emails to venues. Replies are too long and sometimes promise dates the team has not confirmed. She keeps a test set of five real venue requests and changes one thing per round:

Round	One change	Result on the 5 test requests
0	Original prompt	5 too long, 2 promise unconfirmed dates
1	Add "under 120 words" and a three-part structure	0 too long, 2 promise dates
2	Add "Offer only the dates listed in `<dates>`. If none fit, ask the venue for alternatives."	0 too long, 0 promise dates
3	Switch to a faster model to save time	Same quality, quicker replies; keep it

Because each round changes one thing, she knows which change fixed which problem.

Adjusting from feedback

Keep a test set. Three to five real inputs with a known good result. Rerun them after every change.
Change one thing at a time. Prompt wording, a file, a setting or the model, not all together.
Write down what changed and the result. A short log stops the team repeating failed fixes.
Use reviewer feedback as data. If reviewers keep making the same edit, put that edit into the prompt or project instructions.
Turn a reaction into an instruction. "Too generic" is a reaction; "name the audience and the one action they should take" is an instruction. Ask what the output needed to come out right, and where in the setup that comes from. If you cannot name that part, the next attempt is a guess.
Capture the fix where it belongs. A fix left in one chat is rediscovered next week, by you or a colleague covering for you. Ask whether the same correction will be needed again; if so, sort it: a rule goes in project instructions, reference material in project knowledge, a multi-step procedure in a Skill. Memory is per person and picks things up on a best-effort basis, so it is not the place for a team fix.
Know when to stop. If careful changes still leave the output unreliable, the task may need a person or an escalation to Architects or Developers.

Optimising a working process

Waste	Optimisation	Why it helps
Uploading the same files to every new chat	Put them in a Project	Upload once; the help centre notes cached project content counts less against usage limits when reused
Many short follow-up messages	Combine related questions into one clear message	Fewer rounds of back-and-forth
A long chat carried on for days	Start a new chat from a checked summary	Longer chats use more of your allowance and drift
Re-typing the same procedure	Write it as project instructions or a Skill	Same steps every time, no re-typing
Copying data out of Drive by hand	Use the Google Drive connector, within your permissions	Claude reads the current file directly
The most capable model on simple tasks	A faster, lower-cost model	Same quality for less time and usage
Reviewers fixing the same issue every time	Fix it in the prompt or instructions	Removes the repeat edit at the source
Different people get different results on the same task	A shared Project and Skill	Everyone runs the same setup
Tools and connectors switched on that the task does not use	Turn them off	The help centre lists this among ways to make usage go further

To find friction, run one full cycle and write down every manual step you also did last time: repeated pasting, repeated corrections and variation between people are the three signals. Then measure the metric that matters for that workflow. Time is easiest, but a customer-facing report may care more about consistency and fewer errors. Prove a new setup over a few cycles before everyone relies on it; stop tuning once the metric is good enough.

Rules that decide exam answers

Find the cause before changing things. The right answer traces the symptom to a cause; switching models or tools at random is the tempting wrong option.
Change one thing at a time and retest. Several changes at once hide which one worked.
Target the bottleneck you measure. Choose the change that removes the cause of the delay or error you track; a faster model does not help if the time goes on fixing inconsistent formats.
Restart drifting chats from a checked summary. Pushing on in an overloaded chat repeats the problem.
Optimise without lowering quality. A faster model or shorter process is only right if results on the test set stay the same.

Where it appears in the exam

Domain 7 carries 10% of the exam, around 6 of the 60 items. Questions describe a symptom, such as generic replies, invented figures, ignored instructions or a process that takes too long, and ask for the most likely cause, the best next change, or the best way to make the process more efficient.

Two sample questions

These are original Timo practice questions. They are not official exam questions.

Question 1

A recruitment coordinator's prompt for interview invitation emails gives poor results. She changed the prompt wording, the model and the attached template all at once. Results improved for some roles and got worse for others. What should she do next?

A. Return to the last good version and change one element at a time, testing on the same requests.B. Keep all three changes and add more instructions to the prompt for the roles that got worse.C. Switch permanently to the most capable model so that every invitation email gets the strongest reasoning.D. Ask Claude to score each email from one to ten and keep only the emails that score eight or above.

Answer: A. Changing one element at a time on a fixed test set shows which change helped and which hurt. B piles more changes on an unknown cause, C assumes the model is the problem without evidence, and D relies on Claude grading its own output.

Question 2

An analyst asks Claude in a chat to "pull this week's numbers from the shared sales sheet in Drive". Claude replies with figures and says it has "checked the sheet", but none match the real file. The Google Drive connector is not switched on for this chat. What is the most likely cause?

A. The selected model is too small to read spreadsheets accurately, so a larger one is needed.B. Claude read an older version of the sheet that Google Drive had cached from last week.C. Claude never had access to the file and made up plausible figures; turn on the connector.D. The prompt was too short, so Claude guessed which of the sales sheets the analyst meant.

Answer: C. Without the connector, Claude could not open the sheet, and the help centre notes Claude can wrongly claim to have used tools it does not have. A and D blame the model or prompt for a missing connection, and B assumes access that did not exist.

Build exercise

Pick a prompt or Project that gives weak results. Write down the symptom and work through the five causes in order.
Build a test set of five real inputs with known good results. Run your current setup and record the score.
Make one change at a time for three rounds, logging the result of each as in the worked example.
List three wastes in a Claude process you use and apply one optimisation from the table. Rerun the test set to confirm quality held.

Practise this topic

Claude Certified Associate practice exam: free, 20 questions, no sign-up
CCAO-F study guide: all topics
Previous topic: Governance, Risk and Responsible Use

Sources

Claude Certified Associate Foundations Exam Guide, version 1.0, effective July 2026 (Anthropic), domain 7: Troubleshooting and Optimization
Claude Help Center: Usage limit best practices
Claude Help Center: How do usage and length limits work?
Claude Help Center: Claude is producing links that don't work and falsely claiming that it has sent emails or produced external documents
Claude Help Center: Claude is providing incorrect or misleading responses. What's going on?