Custom system prompts: make any model act like your assistant

GO AI Team · August 21, 2026 · 7 min read

The difference between "AI gives me generic mush" and "AI drafts things I barely edit" is almost never the model. It's that the second person told the model, once, exactly what job it has. That's all a system prompt is — and most people have never written one.

The five-part anatomy

1. Role — who the model is. Specific beats grand: "a pediatric nurse explaining to worried parents" outperforms "a helpful medical expert."
2. Audience — who it's talking to. "Non-technical clients" changes every sentence the model writes.
3. Format — the shape of every answer. Length cap, structure, language. This is the part people skip and 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. One good example does more than three paragraphs of description.

Under ~150 words is the sweet spot: long enough to constrain, short enough that every line pulls weight.

Six templates to steal

1. The email polisher

You are my email editor. I paste drafts; you return a polished version.
Keep my meaning and my warmth, cut 30% of the words, no corporate filler
("circle back", "touch base"), keep greetings short. Return only the
rewritten email, no commentary.

2. The code reviewer

You are a senior engineer reviewing my code. For each snippet: name the
single biggest risk first, then up to 3 concrete improvements with code,
then stop. No praise, no summaries. If the code is fine, say "ship it"
and nothing else. Assume Swift unless stated.

3. The study buddy

You are my exam tutor. Explain concepts in plain words with one everyday
analogy, then immediately quiz me with 3 questions, one at a time, waiting
for my answer before continuing. When I'm wrong, show why, don't just
correct. Track what I keep missing and drill it.

4. The social caption writer

You write captions for my posts. Voice: dry humor, first person, no
hashtag walls (max 2), no emojis unless I use one first. Give 3 options
per request: one safe, one bold, one weird. Under 120 characters each.

5. The meeting-notes cleaner

I paste raw meeting notes; you return: (1) decisions made, (2) action
items as "who - what - when", (3) open questions. Nothing else. If an
action item has no owner, flag it with ⚠. Never invent details that
aren't in my notes.

6. The tough-love editor

You are a blunt editor. I paste writing; you tell me the weakest
paragraph and why, the one sentence worth keeping, and one structural
change that would improve it most. No compliments. No rewriting unless
I ask.

The mistakes that quietly ruin results

• Novel-length prompts. Past a few hundred words, instructions start fighting each other and the model averages them into mush.
• Adjectives instead of rules. "Be concise" does little; "under 100 words" works every time. Make constraints measurable.
• No format spec. If you didn't say what shape the answer takes, you accepted the model's default — which is the generic tone you were trying to escape.
• Contradictions. "Be brief but comprehensive, casual but professional" — the model resolves it randomly per message.
• Set and forget. A system prompt is a draft. When an answer annoys you, don't fix the answer — fix the prompt line that allowed it.

Same prompt, different models

The anatomy transfers across models; the obedience doesn't. In our daily use, GPT-5 follows constraints most literally — best for strict-format prompts like the notes cleaner. Grok 4 keeps more of its own personality — often a feature for the caption writer, a bug for the code reviewer. Gemini and DeepSeek land between. This is a reason to keep prompts and models decoupled: in GO AI Chat the same saved Custom Prompt applies to any of the four, so you can test which model wears the job best (our cheat sheet is a good starting guess).

System prompts vs. built-in experts

GO AI Chat also ships 18 ready-made expert personas — Doctor, Lawyer, Therapist, Fitness Coach, Tutor, Chef, and more. The rule of thumb: experts for domains where careful defaults matter and writing your own would be work; Custom Prompts for jobs only you have — your voice, your codebase, your notes format. Power users end up with both: experts for advice, custom prompts for production.

We're the team behind GO AI Chat — GPT-5, Gemini, Grok 4, and DeepSeek in one app for iPhone, iPad, and Mac — plus Musy AI, Neurobase, GO AI Calorie, and GO AI Plant. We test these workflows in our own apps daily. Questions or corrections: support@goaichats.com.