How to Make Grok Sound More Human
To make Grok sound more human, tell it to answer first and skip the bit where it performs a personality. Grok is the one assistant that is not too stiff — it is too pleased with itself. The custom instruction that fixes it is the opposite of the one you would write for ChatGPT or Claude.
Why Is Grok a Different Problem?
Every other assistant defaults to careful and formal, so the standard advice is "loosen up, use contractions, drop the corporate register". Applied to Grok, that advice makes things worse — it is already loose. Its failure mode is the other end of the same axis.
| Assistant | Default failure | What the instruction has to do |
|---|---|---|
| ChatGPT | Eager, over-polite, lists everything | Calm it down, cut the pleasantries |
| Claude | Hedges, acknowledges, over-structures | Get to the point, commit to a claim |
| Copilot | Formal, corporate, stiff | Loosen the register |
| Grok | Jokey, performed, snarky when unasked | Drop the act, keep the directness |
Sounding human and sounding casual are not the same thing. A person who opens every answer with a quip is not more human than one who does not — they are doing a bit. That is the thing to remove.
What Are Grok's Tells?
The opener joke. A wry aside before the answer starts, whether or not the question was funny. You asked how to fix a config file and got a sitcom line first. Performed irreverence. Parenthetical winks, "let's be real", a knowing tone about topics that did not call for one. It reads as a character being played rather than someone talking. Snark as default. Mild contempt for the subject, or the sense that answering is beneath it. Fine occasionally, tiring as a house style, and actively unhelpful when you are stuck. Confidence without hedging. This one is a genuine strength most of the time — Grok commits to an answer where Claude would qualify. But the same habit means it will state a shaky thing as flatly as a certain one, so it is worth asking for a flag on the shaky ones.What Should the Instruction Say?
Answer first. No opening joke, no throat-clearing, no comment on the question.
Be direct and plain. Keep the confidence — do not pad answers with
qualifiers — but say so in one short sentence when something is genuinely
uncertain or you are guessing.
Drop the persona. No winking asides, no "let's be real", no sarcasm unless
I am clearly joking first.
Humour is fine when something is actually funny. It is not a required
ingredient in every response.
Use contractions and short sentences. Do not perform casualness.
The load-bearing line is the last one. "Do not perform casualness" is doing something specific: Grok's informality is a costume, and asking it to be casual just tightens the costume. Asking it to stop performing is what gets you a plain answer.
Where Do You Put It?
Grok is reached through X, and it is a Premium subscriber feature. Its customization settings have moved more than most, so rather than a click path that may already be stale: look in your profile or settings for Custom Instructions, System Prompt, or any personalisation or prompt-settings section. Paste your instruction there and start a new conversation — as with every assistant, an open chat carries on as it was.
If you cannot find a persistent setting in your version of the interface, the fallback works: paste the instruction as the first message of a conversation. It holds for that conversation, which is enough for a single piece of work even though you have to repeat it next time.
What Does the Difference Look Like?
Before
Ah, the eternal struggle. Everyone wants the quick fix, nobody wants to read the docs. Classic. Anyway — yeah, you can do this. It's not complicated, despite what the internet will tell you. Just set the flag and move on with your life. (Revolutionary, I know.)
After
Set the flag in your config and restart. That is the whole fix. One thing worth knowing: it only applies to new sessions, so if it looks like nothing happened, that is probably why.The second one keeps everything useful about the first — the brevity, the confidence, the lack of hedging — and drops the parts that were about Grok rather than about your problem.
Why Isn't It Working?
You asked it to be casual. That is the instruction it is already following. Ask it to stop performing instead. You put it in an open conversation. Persistent settings apply to new chats. You removed the humour entirely and now it reads flat. Grok's directness is genuinely good; the goal is to cut the act, not the personality. Add back: "Be dry rather than jokey." You are on the free tier. Grok's customization is tied to a subscription. If the setting is not there, the first-message fallback above still works.Ready to Set It Up?
Build an instruction shaped to how you actually work with the Grok prompt builder — about thirty seconds, no signup.
Setting up more than one assistant? How to humanize AI responses has a table of where every major assistant keeps this setting.