@hackproduct9: "Write a better prompt" is not advice. These 12 patterns are what people actually mean when they say it. Wrong question: "what's the magic wording?" Right question: "which pattern does this task need?" 12 patterns, one line each 👇 🟢 EXAMPLES 01 Zero-shot — just ask, no examples 02 One-shot — one example sets the shape 03 Few-shot — 3-5 examples and it copies the pattern 🔵 STEERING 04 Chain of thought — make it show working before the answer 05 Role prompting — a job title picks the vocabulary 06 Structured output — JSON your code parses, not prose 🟡 SHAPE 07 XML tags — wrap sections so it sees the boundary 08 Constraints — length, tone, and what to leave out 09 Context — choose what enters the window, drop the rest 🟣 SCALE 10 Templates — one prompt, many inputs, versioned 11 Caching — reuse the stable prefix, pay for less 12 Self-reflection — draft, critique, revise 👀 Watch cards 01-03. Same machine, only the example count changes, and the reliability bar climbs 42% → 68% → 91%. Look where the jump is: going from zero to ONE example buys you more than going from one to three. Most people skip straight to writing five examples when one good one would have done it — and every extra example is prefix you pay for on every single call. Which is why card 11 pairs with it: keep that prefix byte-identical and your billed tokens drop to ~34%. Change one character and you pay full price again. The rule: pick the pattern for the failure you have, not the prompt you wish you'd written. Most people tune wording for hours when the task needed structure. Send this to whoever is on their ninth prompt rewrite. 📸 Screenshot it — 12 patterns, one card. Follow @hackproduct — scary AI concepts, made shippable. ⚡ . . #promptengineering #LLM #AIengineering #GenAI #contextengineering

HackProduct

Open In TikTok:

Region: US

Saturday 08 August 2026 21:31:32 GMT