How to Make Claude Sound More Human
To make Claude sound more human, save a custom instruction telling it to drop the opening acknowledgement, stop hedging every claim, and answer in prose rather than bullet points unless the answer is genuinely a list. Claude's stiffness is not general politeness — it is three specific habits, and each one has a specific instruction that turns it off.
This guide covers what those habits are, why Claude in particular has them, and exactly what to write to stop each one. If you only want the settings path and a prompt to paste, the Claude custom instructions guide is the shorter route.
Why Does Claude Sound Different From ChatGPT?
They fail in different directions, which matters because generic "make it sound human" advice fixes whichever one the writer happened to be using.
| Habit | How Claude does it | How ChatGPT does it |
|---|---|---|
| Opening | "That's a great question. Let me think about this." | "Certainly! I'd be happy to help." |
| Uncertainty | Hedges genuinely simple claims, at length | States things flatly, sometimes over-confidently |
| Structure | Bullets and headers even for a two-sentence answer | Numbered lists for everything |
| Length | Adds caveats and context you did not ask for | Adds a summary of what it just said |
| Corrections | Over-apologises, then re-explains from scratch | Apologises briefly, moves on |
The practical consequence: an instruction written to de-formalise ChatGPT ("be less enthusiastic", "skip the pleasantries") barely touches Claude, because Claude was not being enthusiastic. It was being careful.
What Are Claude's Three Tells?
The acknowledgement
Claude often spends its first sentence telling you that it heard you. "That's a really interesting question." "You raise a good point." Nobody talking to a colleague does this. It reads as filler because it is filler — the answer starts in sentence two.
The hedge
Claude qualifies. Sometimes that is exactly right, because the honest answer really is "it depends". But it also hedges things that are not in doubt, and stacked hedges read as evasion: "generally speaking, in most cases, it's often considered..." Three qualifiers on one uncontested claim is a tell.
The structure reflex
Ask Claude something with two parts and you will frequently get two headers, four bullets and a bold label on each. Structure is genuinely useful for a comparison or a sequence. Applied to a paragraph's worth of thought, it turns a conversation into a briefing document.
What Should the Instruction Actually Say?
Name the habit rather than the mood. "Be more casual" is vague enough that Claude will comply cosmetically — contractions, shorter words — while keeping the acknowledgement, the hedge and the bullets. Instructions that name a behaviour get followed literally, which is what you want here.
Write the way a knowledgeable colleague talks.
Start with the answer. Do not open by acknowledging the question, restating it,
or telling me it is interesting.
Be direct about things you are confident in. Save hedging for claims that are
genuinely uncertain, and hedge once rather than three times.
Default to prose. Use bullets or headers only when the content is genuinely a
list or a comparison — not to organise a single idea.
Use contractions. Vary sentence length. Let some sentences be short.
Do not summarise what you just said at the end of a response.
If I am wrong about something, say so plainly and briefly. No extended apology.
Paste that as-is or use it as a starting point. Every line targets one of the habits above, which is why it is worded as prohibitions rather than adjectives.
Where Do You Put It?
Claude keeps custom instructions in two places, and the difference catches people out.
| Scope | Applies to | Use it for |
|---|---|---|
| Account-level preferences | Every new conversation | How you want Claude to write, generally |
| Project instructions | Chats inside that one project | Instructions specific to a body of work |
Project instructions are the ones most people find first, and they are also the ones that surprise people: a prompt saved inside a project does nothing outside it. If you have set custom instructions once and they seem to have stopped working, check whether you are in the project you saved them to.
To set project instructions: open Projects, select or create a project, choose Edit project, paste your instruction, and save. Then start a new conversation — an open chat keeps behaving the way it already was.
What Does the Difference Look Like?
Before
That's a really interesting question, and I appreciate you asking it. There are several factors to consider here. Key considerations: - Timing: Generally speaking, it's often best to consider the timing carefully - Audience: In most cases, you'll want to think about who's reading - Tone: This can vary depending on context I hope this helps! Let me know if you'd like me to elaborate on any of these points.
After
Send it Tuesday morning. Most people clear their inbox Monday and stop reading carefully by Thursday, so mid-week early is the sweet spot. The bigger question is who it's for. If it's going to the whole team, cut the second paragraph — that detail only matters to the two people already in the thread.Same information. The second one sounds like someone who has an opinion.
Why Isn't It Working?
You changed it mid-conversation. Custom instructions apply to new chats. The one you have open will not change. You saved it to a project and left the project. Project instructions are scoped to that project only. This is the most common cause by a distance. Your instruction described a mood, not a behaviour. "Be more human" and "sound natural" are the two weakest instructions you can write. Claude will comply in ways that do not change the tells. Name the habit. You asked for too much at once. Fifteen rules compete with each other, and the ones near the bottom lose. Start with the three habits above and add only what you actually miss. The task genuinely warrants structure. If you asked for a comparison of six things, bullets are the right answer and Claude is being sensible. The instruction is meant to stop reflexive structure, not all structure.Does This Change What Claude Can Do?
No. These instructions change how Claude writes, not what it knows or how carefully it reasons. That distinction matters because the hedging habit and genuine epistemic caution look similar from the outside and are not the same thing. Telling Claude to stop stacking qualifiers does not tell it to be more confident than it should be — if the honest answer is uncertain, you still want it to say so, once.
If you find responses have become too blunt, add a single line: "When something genuinely is uncertain, say so plainly in one sentence." That restores the useful half of the habit without the padding.
Ready to Set It Up?
Build a prompt tuned to how you actually write, in about thirty seconds, with the Claude prompt builder. Or take the copy-paste route in the Claude custom instructions guide.
Using more than one assistant? How to humanize AI responses covers where every major assistant keeps this setting, in one table.