Study notes · 3.3% of the exam

Identify, diagnose, and resolve issues with underperforming prompts or poor outputs

Work out why a prompt or configuration is producing poor output by checking, in order, the prompt itself, the context and knowledge available, the request size, the conversation length, and where the chat actually lives.

Key points

  1. 1

    Generic output almost always means a generic prompt: no audience, no purpose, no required content, no length or format. The fix is to state who the output is for, what it must cover, and how it should look.

  2. 2

    Identical outputs across different inputs (a shared 'generator' prompt) mean missing context and missing examples. Add the specific inputs each run and one or two examples of what good looks like.

  3. 3

    Vague quality adjectives ('compelling', 'specific', 'creative') and capitalised 'MUST' do not supply missing information; they are the tempting non-fix.

  4. 4

    Wrong format: if you needed a table with particular columns, a bullet list, or a fixed structure, say so explicitly; regenerating and hoping is not a diagnosis.

  5. 5

    Overloaded requests (five tasks in one message) produce shallow results on all of them. Split into sequential requests and check each before continuing.

  6. 6

    Stale Project knowledge produces confident, outdated answers. When a Project consistently quotes old prices, procedures or policies, check the knowledge files before touching instructions or model; replace superseded files and remove the old ones.

  7. 7

    Project instructions and knowledge apply only to chats started inside the Project. A chat opened from the main page with a manual attachment has none of that configuration, however the instructions are worded.

  8. 8

    Very long conversations degrade: contradictions, forgotten decisions and slower replies. Length limits are set by the context window per conversation; the remedy is a fresh chat that opens with a written summary of decisions and constraints.

  9. 9

    Length limits (how long one conversation can be) are different from usage limits (how much you can use Claude across all chats in a period); a plan upgrade does not fix a diluted conversation.

  10. 10

    Switching to a bigger model is rarely the diagnosis; a more capable model still cannot know facts that are absent from the prompt or the knowledge.

  11. 11

    'Remember everything above' and self-check instructions are not remedies for context dilution or calculation errors; structural fixes (restart, split, verify) are.

Test yourself on Identify, diagnose, and resolve issues with underperforming prompts or poor outputs

Ten questions, with the answer and explanation after each one.