# Custom system prompts: make any model act like your assistant

> **TL;DR:** A **system prompt** is a standing job description the model reads before every message: role, audience, format, constraints, one example. Write those five parts, stay under ~150 words, and any model — GPT-5, Gemini, Grok 4, DeepSeek — becomes *your* assistant. In [GO AI Chat](https://goaichat.app/chat), save them once in the Custom Prompts library and reuse across all four models.

By the GO AI Team · August 21, 2026

## Five-part anatomy

1. **Role** — specific beats grand: "a pediatric nurse explaining to worried parents" > "a helpful medical expert."
2. **Audience** — "non-technical clients" changes every sentence.
3. **Format** — length cap, structure, language. The part people skip, then blame the model.
4. **Constraints** — the "never" list: "Never use buzzwords. Never invent numbers. If unsure, say so."
5. **One example** — a single sample answer in your voice; worth three paragraphs of description.

## Six templates

1. **Email polisher:** "You are my email editor. I paste drafts; you return a polished version. Keep my meaning and warmth, cut 30% of words, no corporate filler, short greetings. Return only the rewritten email."
2. **Code reviewer:** "Senior engineer reviewing my code. Per snippet: biggest risk first, then ≤3 concrete improvements with code, then stop. No praise. If fine, say 'ship it'. Assume Swift."
3. **Study buddy:** "Exam tutor. Plain-words explanations with one everyday analogy, then quiz me 3 questions one at a time, waiting for my answer. When I'm wrong, show why. Track and drill what I keep missing."
4. **Caption writer:** "Captions for my posts. Dry humor, first person, max 2 hashtags, no emojis unless I use one. 3 options per request: safe, bold, weird. Under 120 characters."
5. **Meeting-notes cleaner:** "I paste raw notes; you return (1) decisions, (2) action items as who–what–when, (3) open questions. Nothing else. Flag ownerless items ⚠. Never invent details."
6. **Tough-love editor:** "Blunt editor. Tell me the weakest paragraph and why, the one sentence worth keeping, one structural change. No compliments. No rewriting unless asked."

## Mistakes that ruin results

Novel-length prompts (instructions fight, model averages to mush) · adjectives instead of measurable rules ("be concise" vs "under 100 words") · no format spec (you accepted the generic default) · contradictions ("brief but comprehensive") · set-and-forget (when an answer annoys you, fix the prompt line that allowed it).

## Same prompt, different models

Anatomy transfers; obedience varies. In our daily use: **GPT-5** most literal (best for strict formats), **Grok 4** keeps most personality (feature for captions, bug for code review), **Gemini**/**DeepSeek** in between. In GO AI Chat one saved Custom Prompt applies to any of the four — test which model wears the job best. Starting guesses: [model cheat sheet](https://goaichat.app/blog/which-ai-model-for-which-task).

**Custom Prompts vs the 18 built-in experts:** experts (Doctor, Lawyer, Therapist, Coach, Tutor, Chef…) for domains where careful defaults matter; Custom Prompts for jobs only you have — your voice, your codebase, your format.

## FAQ

- **System prompt?** Standing instructions read before every message, whole-conversation scope.
- **Vs regular prompt?** Request vs job description.
- **Same on every model?** Idea yes, obedience no — save once, test per model.
- **Saving in GO AI Chat?** Custom Prompts library; works with all four models. [App Store](https://apps.apple.com/app/id6742322421)

---

More: [GO AI Blog](https://goaichat.app/blog) · [GO AI Chat](https://goaichat.app/chat) · support@goaichats.com

© 2026 GO AI Chat LLC