Prompt Engineer question bank
100 questions. Filter by level and skill, track what you've solved, and see how many people get each one right.
- Improving a vague prompt
- Few-shot prompting
- Separating instructions from data
- Giving the model a role
- Getting a consistent format
- Zero-shot prompting
- Reducing made-up answers
- Testing a prompt properly
- What a token is
- What fills the context window
- Raising the temperature
- What top-p controls
- Reply cut off mid-sentence
- Stop sequences
- The model forgot the last message
- Questions about last week
- Using XML-style tags
- Prefilling the response
- Chain-of-thought prompting
- What RAG means
- What embeddings are for
- How function calling works
- What a hallucination is
- Jailbreak or prompt injection
- Customer data in prompts
- What a golden set is
- Keeping track of prompt changes
- Quote first, then answer
- Message roles in chat APIs
- Same text, more tokens
- What you pay for
- Temperature 0 still varies
- Making replies faster
- Asterisks in text messages
- Examples that bias the label
- When to ask for reasoning
- One huge prompt that fails
- Negative instructions
- Where to put the question
- Judging summary quality
- Where to put fixed rules
- Settings for a ticket classifier
- JSON that sometimes breaks
- Handling invalid model output
- Choosing a chunk size
- Search misses product codes
- Right chunk at rank 15
- Answers beyond the documents
- Checkable citations
- Writing a tool description
- When a tool call fails
- An agent that never stops
- Getting prompt cache hits
- Labelling a million reviews
- Big model for every request
- The judge prefers answer A
- A small prompt fix ships
- Which metric for auto-refunds
- Sampling several answers
- Redacting before the API call
- Checking replies before users see them
- Debugging a contradictory prompt
- Every reply sounds like example 1
- Facts buried mid-context
- Parsing an answer after reasoning
- Picking examples per input
- Vague 1-10 judge scores
- Answers too basic for experts
- Instructions hidden in user content
- Perfect on 20 examples
- Can we trust the LLM judge
- Switching to a newer model
- Prompting a reasoning model
- Adding 30 more examples
- Invented values in extraction
- Agent picks the wrong tool
- Injection through a web page
- Leaking data through an image link
- Prompt cache hit rate near zero
- Judge and candidate from one family
- A suspicious score jump
- Is a 2-point gain real?
- Follow-up questions fail in RAG
- Where is the RAG bug?
- Strict JSON hurt accuracy
- Forced into the wrong label
- An agent that deletes data
- Hundreds of fake dialogues
- Agent with 120 tools
- Real sources, wrong claims
- Longer answers keep winning
- Tiny wording, big swing
- Errors flowing down a chain
- Scores change between runs
- Reasoning model bill surprise
- The model agrees with everything
- API key in the system prompt
- Fresh facts plus a fixed style
- Long agent run fills context
- Evals green, users unhappy