Day 02·Practice and review·2 min read
Cheat sheet
How scenario questions are graded, the rules of thumb that catch distractors, the symptom-to-fix table, and the effects worth memorising.
By the end
Use the symptom-to-fix table and the grading rules to pick the answer a scenario question is looking for.
These questions reward hedged answers with a real mechanism behind them, and punish absolutes. Match each option to a specific symptom in the scenario before you pick it.
Rules of thumb
- "May", "some" and "occasionally" outcomes are usually right when you can name the mechanism.
- "Consistently", "always", "significantly", "entirely" and "guarantee" usually mark a distractor.
- Options that remove information or ignore the user's request are wrong.
- Options that trade the real goal for something else, such as complex language "for credibility", are wrong.
- Two symptoms in the scenario usually means at least two correct options, one per symptom.
Symptom to fix
| Symptom | First fix | Also works |
|---|---|---|
| Generic, uninspired output | Concrete instructions plus diverse examples | Higher temperature |
| Repetitive, similar outputs | Higher temperature or top-p, varied examples | Generate many, then filter |
| Breaks a length or format rule | Examples that obey it plus an explicit limit | Schema, validate and regenerate |
| Vague answers to ambiguous input | Instruct it to ask clarifying questions, with examples | CoT to spot what is missing |
| Misses key items in summaries | Name the items, then extract-then-summarize | Examples of good summaries |
| Too much jargon for the audience | Name the audience and the jargon rule | Plain-language examples, two-step rewrite |
| Invents facts | RAG with "answer only from context" | Lower temperature, require citations |
| Wrong on multi-step reasoning | Chain-of-thought | Self-consistency |
| Needs live data or actions | Tool use with a ReAct loop | Prompt chaining |
Effects to memorise
- Higher temperature: more variety, less rule-following, occasional odd or abstract output. Never more repetitive.
- Lower max tokens: truncation, not clarity.
- Negative instructions: partial compliance at best.
- Few-shot examples: the strongest lever, with an anchoring risk.
Test yourself
8 multiple-choice questions and 15 flashcards on this lesson, shuffled every time, with the reason behind every answer.