CCAO-F Prompting and Context Engineering Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Ask Claude to summarize the entire manual first, then answer the question based only on that summary.

  Why it's wrong here

  Summarizing a 90,000-token manual before answering introduces information loss and a second layer of potential error. Details needed for a specific question may be dropped or distorted in the summary, and the model then answers from a lossy intermediate representation. Direct grounding in the full source, combined with clear instructions, is more reliable than relying on a self-generated summary.

• ✓

  Instruct Claude to answer only from the provided manual and to say 'I don't know' if the answer is not present.

  Why this is correct

  Explicitly grounding the answer in the provided manual and allowing a 'I don't know' response reduces hallucination and encourages the model to rely on the source. This is especially important with long documents where the model might otherwise fill gaps with general knowledge. It also gives users a clear signal when the manual lacks the needed information.

• ✓

  Place the long manual content before the user's question and the specific instructions, so the model reads the source material first.

  Why this is correct

  Putting long source documents before the question and instructions aligns with how attention and instruction-following tend to work in long-context prompts. The model processes the manual first and then applies the specific question and constraints, improving the chance that relevant sections influence the answer. This ordering also helps the instructions remain salient at the end of the prompt.

• ✗

  Increase the temperature to 1.0 so Claude considers a wider range of interpretations from the manual.

  Why it's wrong here

  Higher temperature increases randomness and diversity, which is counterproductive for factual question answering over a fixed manual. It raises the risk of inconsistent or invented answers rather than improving accuracy. For grounded retrieval-style tasks, lower temperature is generally preferred to keep outputs focused and reproducible, so this technique works against the developer's goals.

• ✗

  Split the manual into random fragments and include only a few fragments per request to keep the prompt short.

  Why it's wrong here

  Randomly including fragments is essentially a poor retrieval strategy that can omit the very section containing the answer. Keeping prompts short by random sampling sacrifices completeness and consistency, leading to unreliable answers. If context limits are a concern, a deliberate retrieval method that selects relevant sections is needed, not random fragment selection.