> prompting basics
Most "ChatGPT gave me a bad answer" problems are prompt problems. The techniques here — naming a role and format, separating instructions from content, showing examples, and asking for reasoning — come straight from OpenAI's own prompt-engineering guidance and transfer to almost any request. Specific instructions beat clever ones.
// prompting basics
6 entries> Act as a senior UX writer. Rewrite the onboarding screen below for first-time users: plain, encouraging, max 40 words, no jargon. [paste text]> Summarize the text below as a bullet-point list of the most important points. Text: """[paste text]"""> Extract the important entities from the text below. Desired format: Company names: <comma-separated> People names: <comma-separated> Specific topics: <comma-separated> General themes: <comma-separated>. Text: [paste text]> Extract keywords from each text. Text 1: Stripe provides APIs for payment processing. Keywords 1: Stripe, payment processing, APIs ## Text 2: OpenAI trains language models available through an API. Keywords 2: OpenAI, language models, API ## Text 3: [paste text] Keywords 3:> Work through this step by step before answering, then give the final answer on its own line: [paste problem]> You are a support agent. Diagnose the customer's login problem and suggest a fix. Do not ask for personal information such as username or password; instead, point them to the reset guide at [help URL]. Customer: [paste message]// faq
Why does ChatGPT give generic or off-target answers?
Usually because the prompt is under-specified. Name the role you want it to take, the exact task, the output format, and constraints like length and audience. A concrete "as a UX writer, rewrite this in under 40 words" outperforms "make this better" every time.
What is few-shot prompting?
Showing two or three example input-to-output pairs before your real input so the model copies the pattern. Use it when a plain instruction gives inconsistent results. Start zero-shot with no examples and add them only if needed, since examples use context and can bias the output.
Should I tell ChatGPT what not to do?
Prefer telling it what to do instead. Negative rules ("don't ask for personal info") leave it without a next action; positive redirection ("refer them to the reset page") gives it a path. OpenAI's best-practices guide recommends this directly.
Does asking ChatGPT to think step by step actually help?
For math, logic, and multi-part tasks, yes — reasoning in the open catches errors a straight-to-the-answer reply hides. On newer reasoning models much of this happens internally, but explicitly asking still helps on hard prompts. Add "then give just the final answer" if you do not want the working.