Ask HN: What is your ChatGPT customization prompt?

Hacker News Discussion ↗

Power users are increasingly crafting elaborate “custom instructions” for ChatGPT and similar models, trying to make them more terse, less moralizing, and better aligned with their coding or research workflows. Many describe prompts that push the model to act as a domain expert, think step‑by‑step, avoid boilerplate safety disclaimers, and prioritize concise code over explanations, while others prefer defaults or minimal tweaks and treat the whole exercise as borderline cargo cult. A recurring theme is the trade‑off between brevity and quality of reasoning, and uncertainty over how much these prompt rituals actually change model behavior versus simply shifting where the failures and hallucinations appear.

Overall Theme

  • Thread collects people’s “custom instructions” / system prompts for ChatGPT and similar LLMs.
  • Main goals: reduce verbosity, improve code quality, suppress moralizing, and shape personality or style.

Common Custom Instructions

  • “Be terse / concise / no yapping / no essays” is by far the most common request.
  • Many ask for:
    • Direct answers first, explanations later.
    • Code-first responses, often with specific language, style, or stack (e.g., ESM imports, async/await, Tailwind, Elixir by default, no semicolons in JS).
    • No restating the question, no apologies, no disclaimers, no “I’m an AI…”.
    • Avoid numbered lists; prefer summaries or prose.
  • Several prompts define roles: “expert in X”, software architect, polymath, medical assistant, etc.

Terseness vs Verbosity

  • Split in preferences:
    • Some insist brevity boosts usability and reduces boilerplate, especially on slow models.
    • Others argue longer, step-by-step “chain-of-thought” answers improve quality and reliability.
  • A hybrid approach appears popular: detailed internal reasoning but short summary at the end, or verbosity controlled via a flag (e.g., V=0–5, or keywords like “vv”).

Reasoning, Computation & Prompt Theory

  • Repeated idea: each extra token is more “computation,” which may improve reasoning.
  • Some instruct models explicitly to:
    • State assumptions.
    • Break problems into steps.
    • Provide multiple solutions or perspectives.
    • Self-check and correct earlier answers (with mixed success).
  • Discussion of research showing “think step by step” / “take a deep breath” can help; not everyone is convinced more tokens always help.

Ethics, Safety & Tone

  • Many users explicitly try to disable:
    • Moral lectures, safety disclaimers, or “political correctness”.
    • Suggestions to seek professionals or other sources.
  • Several stress neutrality and fact-focus; corrections are desired when facts are wrong.
  • Some find the constant safety framing akin to “coffee is hot” warnings; others note it originates from past chatbot failures and PR concerns.

Humor, Abuse & Anthropomorphism

  • Numerous playful or adversarial prompts: pirate talk, Ali G, snark, calling PowerShell “StupidShell,” threats of “death,” tips for saving kittens, being in love with the user, etc.
  • Some commenters find this fun; others find it depressing or “tribal/ritualistic,” likening it to incantations before a black box.

Skepticism & Practical Limits

  • Several report that models often ignore instructions (especially brevity, partial-code-only, or “never say X”).
  • Some doubt that long, intricate meta-prompts help much beyond what defaults already provide.
  • Others prefer no customization at all, relying on conversational steering and follow-up questions instead.