Domain 7 · lesson
Troubleshooting and Optimization
Learning goal: identify the actual cause of a failure and improve the workflow without discarding required quality or risk controls. A useful diagnostic order is evidence/permissions -> instructions -> context -> execution -> validation -> capability. This is a study heuristic, not an official rigid checklist. S07 S09 S18 S21
Objective 7.1
Diagnose poor prompts and outputs
Core content
Describe the symptom observably: invented facts, missing section, stale policy, unreadable file, quota exhaustion, wrong format, or failed connector action. Form a causal hypothesis and use the smallest test that distinguishes it from alternatives. S10 S11 S18 S21
Applied example
If a summary contradicts a new policy, first verify which file was actually available. If the correct file was retrieved but an exception was missed, inspect extraction and context. If only tone is wrong, examine instructions and examples. Changing file, model, prompt, and configuration at the same time destroys diagnostic information.
Exam decision rule
The best next action often reduces uncertainty about the cause rather than adding more technology.
Objective 7.2
Adjust the approach using feedback and results
Core content
Give specific feedback, preserve what already works, and verify the correction. Self-correction can help when the model has the needed evidence; it cannot manufacture a missing source or authorization. Define a stopping condition for retries. S08 S09 S10
Applied example
A table's rows sum to 145 while the reported total is 150. Point to the mismatch and verify with an independent calculation. If a required cost is absent from the data, do not silently estimate it merely to force the total to close.
Exam decision rule
Fixing the known example does not prove the whole problem is solved. Rerun boundary cases and an independent sample.
Objective 7.3
Optimize efficiency and effectiveness
Core content
Optimize the whole task while preserving quality and risk thresholds. Levers include removing irrelevant context, reusing approved configuration, selecting a more efficient model, reducing redundant stages, and automating objective checks. None should remove necessary evidence or critical controls. S22 S25 S09
Applied example
Cutting generation from three minutes to thirty seconds while adding four minutes of manual correction is not an end-to-end improvement. Removing a redundant rewrite while preserving validation may be.
Exam decision rule
More agents, a longer prompt, or the most expensive model are not quality metrics. Compare total cost, accepted-output rate, latency, review effort, and operational complexity.