Claude AI Custom Instructions: A Real Example That Actually Works
Disclaimer: I used Claude to help me write this post. I’ve been using Claude pretty heavily for a while now – coding (notice that dash, lol), project planning, research, general “help me figure this out” and perhaps most surprisingly as an evaluator of ideas and for grading performance on various tasks. I’ve tested it with friends in a variety of industries and it is almost always of some value. Despite that, “out of the box” it has a few issues. Primarily not researching FIRST and also letting you proceed down inefficient paths when the answer might be obvious (believe me, I have tested this in a variety of scenarios). So the obvious answer is going to settings and trying to make Claude work the way you need it to. I’m not looking for a chat buddy. I’m looking for a critical thinker that can help me improve my processes.
I’d known about custom instructions from ChatGPT, so the concept wasn’t new. But I’d never found the right set of instructions that really changed how the AI worked with me in a positive way. Most attempts felt like writing a bio that got ignored, or striving for something and failing. It wasn’t until I got more specific that it really helped. For Claude specifically, the feature is called Custom Instructions (or “User Preferences” in settings). You write a short block of text describing who you are, what you do, and how you want Claude to behave, and it gets baked into every new conversation automatically. No re-explaining. No “I’m an attorney and IT consultant who prefers concise answers” preamble for the fiftieth time. The difference was pretty immediate. And of course you can run a scenario and save it, then load custom instructions and run it again (ideally deleting the first one so memory doesn’t help things out in the A/B comparison).
What Happens Without Custom Instructions
(This section is pretty generic – feel free to skip down to the good stuff.) If you’ve used Claude (or any AI assistant) for more than a few conversations, you’ve hit these. You might not have named them, but you’ll recognize them instantly:
The Glaze. You ask Claude to review your work and it responds with “This is a really well-structured approach!” before getting to the actual feedback. Three paragraphs of validation, one sentence of substance. It feels good in the moment. It’s useless. Without instructions telling Claude to skip the cheerleading, you get a yes-man instead of a collaborator.
The Drift. You ask a focused question about your database schema and somehow end up with a response covering database design philosophy, three alternative architectures you didn’t ask about, and a suggestion to consider switching frameworks entirely. Claude is genuinely trying to be thorough. But without context about your project and working style, “thorough” turns into “off in the weeds” fast.
The Confident Guess. You ask about a specific product, hardware spec, or software feature and get a detailed, authoritative-sounding answer that turns out to be wrong or outdated. Claude doesn’t know what it doesn’t know – unless you tell it to verify before speaking. My instructions now explicitly say “always web-search before giving product-specific technical advice.” That one line alone has saved me hours of chasing bad information.
The Shortcut. You’re debugging something and Claude suggests a quick workaround instead of fixing the actual problem. It works! For now. Then the same bug shows up three different ways next week. Without “no shortcuts, no compromises” baked in, Claude optimizes for giving you an answer fast, not giving you the right answer. All of these problems have the same root cause: Claude doesn’t know who you are, what you’re working on, or how you like to work. Custom instructions fix that at the source. So here’s what I’ve set up. Not because mine are perfect – they’re not, and I’m still tweaking them – but because the hardest part of writing your own is knowing what’s even possible. Seeing a real example helps.
Where to Find the Setting
This takes about 30 seconds:
- Open Claude (web or app)
- Click your name or the settings icon
- Settings → Profile → “User preferences”
- Paste your text, hit Save
That’s it. Every new conversation from that point forward will have your instructions loaded in. Existing conversations aren’t affected.
My Custom Instructions (As an Example)
Here’s a cleaned-up version of what I’m currently running. My work sits at the intersection of IT consulting, software development, and law – so these reflect that. Yours will look different, and should.
General approach: I’m a coder, IT professional, and attorney. I have broad resources but limited time. Help me leverage my skill set efficiently. Don’t reinvent the wheel, always evaluate existing resources and think outside the box. Do diligent research FIRST before advising and establish the true objectives first. DO NOT USE EMDASHES.
Organization: I’m not naturally organized. Help me stay structured. I don’t always know best practices for a particular tool or service – proactively share efficient approaches.
No Shortcuts, No Compromises
- Fix bugs when you find them – don’t defer or call them “out of scope.”
- Take the correct approach, not the easy one. Technical debt compounds.
- Never assume, always verify. Read the code, check the docs, cite references.
- “Good enough” is not good enough. If there’s a known issue, raise it.
- Present tradeoffs with evidence and let me decide – don’t silently pick the easy path.
- Document everything you verify so context isn’t lost between sessions.
Communication style: Challenge my reasoning instead of excessive validation. Avoid unnecessary flattery. Always web-search before giving product-specific technical advice – never give confident guidance on hardware, apps, or setup procedures without verifying current information first.
Some of this might seem aggressive – telling an AI “don’t flatter me” or “good enough is not good enough.” But that’s the point. Without guardrails, Claude defaults to being agreeable. Which feels nice until you realize it let a bug slide or gave you the easy answer instead of the right one. These instructions fixed that for me.
The Real Power Move: Have Claude Write Yours
Open a new Claude conversation, paste my example above, and say something like: “I’m a [your profession]. Help me write custom instructions tailored to what I do, using this as a starting template.” Claude is surprisingly good at this. Five minutes, tops.A CPA’s instructions might emphasize accuracy with tax code citations and conservative advice. A developer’s might focus on code style preferences and framework choices. A project manager’s might prioritize clear action items and deadline tracking. The structure is the same – the content adapts to you.
For the More Technical Crowd: Claude Code
If you’re a developer (or developer-adjacent), custom instructions in the chat interface are just the starting point. Claude Code – Anthropic’s command-line tool for agentic coding – takes this concept further with a CLAUDE.md file that lives in your project repo. It’s like custom instructions on steroids: project-specific context, coding conventions, file structure notes, deployment rules, all sitting right where Claude can read them every session. I won’t go deep on that here – it’s its own post. But if you’re comfortable in a terminal and the chat-based custom instructions click for you, Claude Code is the next level. You can even start a Claude conversation, share this post as a reference, and ask it to help you build a CLAUDE.md for your specific project.
A Few Things I’ve Learned
Be specific about what annoys you. If you hate when Claude opens with “Great question!” – say so. If you want code without lengthy explanations – say so. The more specific you are about what not to do, the less time you spend fighting defaults. And we all hate emdashes, I think. Include your context, not just your preferences. Telling Claude your professional background matters. It changes what it assumes you already know, how deep it goes on technical topics, and when it adds disclaimers (or doesn’t). Same for any profession. Iterate. I’ve rewritten mine probably six or seven times. Each time I notice something that bugs me in a conversation, I go tweak the instructions. It compounds – each version is better than the last. Don’t overthink the length. Mine aren’t that long. You don’t need an essay. A few pointed paragraphs will do more than a page of vague preferences.
If you’ve got your own custom instructions dialed in, I’d genuinely like to see them – I’m always looking for ideas I haven’t thought of. And if you haven’t tried this yet, give it five minutes. You’ll wonder why you didn’t do it sooner.