Prompt Engineer
Prompt engineering questions
188 questions on Prompt engineering for prompt engineers, easiest first.
Easy 59
- Improving a vague prompteasy
- Few-shot promptingeasy
- Separating instructions from dataeasy
- Giving the model a roleeasy
- Getting a consistent formateasy
- Zero-shot promptingeasy
- The model forgot the last messageeasy
- Using XML-style tagseasy
- Prefilling the responseeasy
- Chain-of-thought promptingeasy
- Keeping track of prompt changeseasy
- Message roles in chat APIseasy
- Same text, more tokenseasy
- Asterisks in text messageseasy
- Tracking broken JSONeasy
- What top_p controlseasy
- What a stop sequence doeseasy
- Reply ends mid-sentenceeasy
- Estimating tokens from wordseasy
- Letting the model say it does not knoweasy
- Asking for a precise lengtheasy
- The model gets the date wrongeasy
- What prompt chaining meanseasy
- Saying why a rule existseasy
- Totals that are slightly offeasy
- Replying in the customer's languageeasy
- Defining what "urgent" meanseasy
- Removing chatty openerseasy
- Prompting for a voice assistanteasy
- One prompt doing five jobseasy
- What presence_penalty doeseasy
- Temperature set to 1.9easy
- Changing both sampling knobseasy
- Asking about last week's newseasy
- Counting the r's in a wordeasy
- Replies sent on WhatsAppeasy
- "Hello! Thanks for contacting us" every turneasy
- SQL with made-up columnseasy
- Code that uses an old APIeasy
- Summaries nobody can useeasy
- Matching the brand voiceeasy
- Brainstorming with temperature 0easy
- Five headlines that all look alikeeasy
- Customers writing in Hinglisheasy
- Writing ₹ amounts the Indian wayeasy
- Getting the same JSON keyseasy
- When to hand off to a humaneasy
- Instructions with a fixed ordereasy
- Why use prompt templateseasy
- Showing what good looks likeeasy
- Links that lead nowhereeasy
- Saying what the output is foreasy
- Emojis in formal emailseasy
- What max_tokens limitseasy
- What a 32k context window coverseasy
- Replies that ignore the customereasy
- Input and output token priceseasy
- Reports with random sectionseasy
- A draft that is almost righteasy
Medium 74
- Examples that bias the labelmedium
- When to ask for reasoningmedium
- Negative instructionsmedium
- Where to put the questionmedium
- Where to put fixed rulesmedium
- Settings for a ticket classifiermedium
- JSON that sometimes breaksmedium
- Handling invalid model outputmedium
- Sampling several answersmedium
- Debugging a contradictory promptmedium
- Every reply sounds like example 1medium
- Parsing an answer after reasoningmedium
- Picking examples per inputmedium
- Answers too basic for expertsmedium
- A classifier that can only say five wordsmedium
- The last example winsmedium
- Label codes versus label wordsmedium
- A greeting with {customer_name}medium
- Template crash on JSON examplemedium
- Markdown in, markdown outmedium
- CRITICAL in capitals everywheremedium
- Long chats near the context limitmedium
- Facts missed in the middlemedium
- Making the model always call a toolmedium
- Labels with stray spellingmedium
- Retrying failed validationmedium
- Is 04/05/2026 April or Maymedium
- Examples as fake chat turnsmedium
- Using an LLM to improve a promptmedium
- Asking the general question firstmedium
- Think, act, observemedium
- Too little detail to answermedium
- Searching with a draft answermedium
- Follow-ups that retrieve nothingmedium
- Brand names getting translatedmedium
- What top_k = 40 meansmedium
- Different answers at temperature 0medium
- A report too long for one replymedium
- Citations you can verifymedium
- Rules a user should not changemedium
- A security team gets refusedmedium
- A bare-minimum dashboardmedium
- Looking up three cities one by onemedium
- Tool output filling the contextmedium
- Order of image and textmedium
- Pulling a table out of a PDFmedium
- Choosing among 200 categoriesmedium
- Examples that are all easymedium
- A prompt with 70 rulesmedium
- Quotes inside JSON stringsmedium
- Sending 500 table rowsmedium
- Arabic that sounds too formalmedium
- Summaries that add factsmedium
- Bad arguments to a toolmedium
- Two ways to get reasoningmedium
- Users seeing the model's workingmedium
- How many examples is enoughmedium
- A persona that driftsmedium
- A useful code-review promptmedium
- Writing background facts firstmedium
- Asking itself sub-questionsmedium
- Tickets with two topicsmedium
- Checking extracted fields quicklymedium
- Showing what not to writemedium
- A shopping chat that loses the cartmedium
- Chatty questions to a keyword indexmedium
- Ranking 30 candidates in one promptmedium
- Translating strings with variablesmedium
- Examples mistaken for real inputmedium
- The model trusts memory over the documentmedium
- Making sure all items are coveredmedium
- Collecting details by chatmedium
- An agent with no finish linemedium
- Schema alone gives odd valuesmedium
Hard 55
- Prompting a reasoning modelhard
- Adding 30 more exampleshard
- Invented values in extractionhard
- Agent picks the wrong toolhard
- Prompt cache hit rate near zerohard
- Where is the RAG bug?hard
- Strict JSON hurt accuracyhard
- Forced into the wrong labelhard
- Tiny wording, big swinghard
- Reasoning model bill surprisehard
- The model agrees with everythinghard
- Prefill that ends with a spacehard
- Savings from a cached prefixhard
- How much room is left for documentshard
- A more honest confidence scorehard
- Summarising a 900-page reporthard
- Rule says one thing, examples anotherhard
- Thinking budget and max_tokenshard
- A rule forgotten in long promptshard
- Solving easy sub-problems firsthard
- Review in the same call or a new onehard
- Planning verification questionshard
- Making a slow reply fasterhard
- Cost of an overnight batch jobhard
- When prompting is not enoughhard
- An expert persona that still errshard
- Exploring several reasoning pathshard
- Denser summaries at the same lengthhard
- Showing JSON while it streamshard
- An order ID nobody gavehard
- A strict 160-character limithard
- Few-shot examples written by an LLMhard
- Examples with random labelshard
- Banning a word with logit_biashard
- An agent losing its placehard
- Old and new policies both retrievedhard
- What a long system prompt costshard
- Caching and time to first tokenhard
- User text that closes your taghard
- Twenty reviews in one callhard
- Sampling code until tests passhard
- Cost of a small-then-large cascadehard
- Per-request tool lists and cachinghard
- Quality drops in long agent sessionshard
- Delegating a messy searchhard
- An LLM that favours option Ahard
- Tips and threats in promptshard
- Optimising prompts from datahard
- Forcing a PAN number formathard
- Names from the examples appearinghard
- Where to put tool usage ruleshard
- How many tokens min_p keepshard
- Top-token probability at T = 0.5hard
- Tokens kept at top_p = 0.8hard
- Field rules inside the schemahard