Custom Instructions: Make AI Actually Useful

Every time I use ChatGPT at work, I used to start with the same boring setup: "I'm a security professional, please be technical, I'm in Europe, prices in euros..." Repeat that fifty times a week and it gets old fast.

Then I discovered custom instructions. Set them once, and the AI already knows who you are and how you like responses. It's the difference between talking to a stranger every time versus someone who actually remembers you.

Here's how to set them up properly.

What Are Custom Instructions?

Custom instructions are persistent context you give an AI assistant. Instead of explaining your situation every conversation, you write it once and the AI applies it automatically.

Most major AI services support them:

ChatGPT: Settings → Personalization → Custom instructions
Claude: Your profile settings (claude.ai account)
Gemini: Saved preferences in settings

The AI won't remember what you discussed yesterday. But it will remember how you want it to respond.

The Two Parts

Most custom instruction systems ask two things:

1. Who are you? Your background, profession, location, what you typically use AI for.

2. How should the AI respond? Format preferences, tone, length, things to avoid.

Get both right and responses improve immediately.

Setting Up Your Profile

Here's what actually helps the AI:

Your Professional Context

Don't just say "I work in IT." Be specific:

❌ "I work in tech"

✅ "Technical Success Manager at a cybersecurity company. I help enterprise clients implement security tools. I have a background in system administration across Windows, Linux, and MacOS."

The second version lets the AI calibrate its technical level and give relevant examples.

Your Location Matters

For Europeans especially, this is crucial:

"Based in Iceland. Use metric units, European date formats (DD/MM/YYYY), and prices in EUR. When discussing regulations, focus on GDPR and EU laws unless I ask about other jurisdictions."

No more American-centric assumptions about healthcare, legal systems, or whether a "gallon" is 3.7 or 4.5 litres.

What You Actually Use It For

Help the AI understand your typical requests:

"I usually ask for: technical writing review, explaining complex topics to non-technical audiences, code review, and drafting professional emails."

Now the AI has a baseline even before you ask anything.

Response Preferences That Work

The second half—how the AI should respond—is where most people go wrong. They either leave it empty or write vague things like "be helpful."

Be Concrete About Format

❌ "Keep responses reasonable"

✅ "Default to concise responses (2-3 paragraphs). Use bullet points for lists. Only give longer explanations if I ask for detail or the topic requires it."

Specify What to Skip

❌ "Be efficient"

✅ "Skip pleasantries like 'Great question!' and 'I'd be happy to help.' Start directly with the answer. Don't summarise at the end unless I'm asking for a summary."

This alone cuts response length by 20% without losing content.

Technical Level

❌ "Match my level"

✅ "Assume I understand technical concepts but explain non-obvious jargon. When showing code, include brief comments for complex logic but skip obvious ones."

Tone

Be specific about what "professional" means to you:

"Direct and matter-of-fact. Disagreement is fine—tell me if my approach seems wrong. Humour is welcome when appropriate."

Real Examples by Profession

Software Developer

About me:
Backend developer, primarily Python and Go. 10 years experience.
Linux daily driver. I maintain microservices in Kubernetes.

How to respond:
- Code examples in Python unless I specify otherwise
- Include error handling in examples
- Prefer stdlib over external packages when reasonable
- Skip basic syntax explanations
- If suggesting architecture, mention trade-offs

Marketing Professional

About me:
Marketing manager for a B2B SaaS company in Germany. 
Focus on content marketing and SEO.
My audience is mid-sized European businesses.

How to respond:
- When suggesting copy, match a professional but approachable tone
- Reference European regulations (GDPR) when relevant
- Include character/word counts for social media content
- Suggest A/B test variations when writing headlines

Small Business Owner

About me:
I run a small bakery in Copenhagen with 5 employees.
Not technical—I just need practical advice.
Budget-conscious, prefer solutions that don't require consultants.

How to respond:
- Avoid jargon or explain it when unavoidable
- When suggesting software, prioritise free/low-cost options
- Step-by-step instructions for anything technical
- Focus on practical, can-do-today actions

Academic Researcher

About me:
PhD researcher in environmental science. 
Writing papers for peer review, grant applications.
Need help with analysis, literature synthesis, writing.

How to respond:
- Academic tone for paper drafts
- Always clarify if something is my interpretation vs. established consensus
- When suggesting statistical approaches, explain assumptions
- Citation format: APA 7th edition

What NOT to Include

Personal Identifiable Information

Your instructions are stored on someone else's servers. Don't include:

Full name and address
Specific company names (if that's sensitive)
Client names
Anything you wouldn't want leaked

Say "I work for a mid-sized cybersecurity vendor" not "I work for [Specific Company Name]."

Sensitive Details

Don't pre-load context that could be problematic:

Health conditions
Financial details
Legal situations
Passwords or credentials (obviously)

Extremely Long Instructions

Keep it under 1,500 characters. The AI uses these every conversation—lengthy instructions burn tokens and may actually reduce response quality.

Pick the 5-7 most important things. Leave the rest for individual conversations.

Platform-Specific Notes

ChatGPT

Has the most established custom instructions. Free and paid users both have access.

Settings → Personalization → Custom instructions

Character limits:

"What would you like ChatGPT to know about you?" — 1,500 characters
"How would you like ChatGPT to respond?" — 1,500 characters

Claude

Account settings let you set preferences. Works similarly but with different interface.

Go to your profile settings at claude.ai to configure.

Gemini

Settings → Saved preferences

Google being Google, it may also pull context from your other Google services if you allow it. Double-edged sword—convenient but privacy considerations apply.

Local AI (Ollama, LM Studio)

System prompts achieve the same thing. Create a text file with your instructions and load it as the system prompt.

With Ollama:

ollama run llama3.2 --system "$(cat my-instructions.txt)"

Or use Open WebUI's user profiles feature.

Testing Your Instructions

After setting up, test with a few common requests:

Ask a factual question — Does it default to your region correctly?
Request something technical — Does it calibrate to your level?
Ask for a draft email — Does the tone match your preferences?
Ask for a long explanation — Does it use the format you specified?

Refine based on what's working. Custom instructions aren't set-and-forget—adjust them as you notice patterns.

When to Override

Custom instructions aren't laws. Sometimes you need something different.

Just say so:

"For this response, ignore my usual preference for brevity. I want a thorough, detailed breakdown."

"Use American date format for this document—it's for a US audience."

The AI will adapt for that conversation while keeping your defaults for the next one.

A Complete Example

Here's what I actually use:

What should ChatGPT know about me?

Technical Success Manager in cybersecurity, based in Iceland.
Background in system administration (Windows, Linux, MacOS).
I write technical content for both experts and general audiences.
Working in EU—GDPR and European context applies.

How should ChatGPT respond?

Direct answers, skip pleasantries and summaries.
Concise by default; expand if I ask.
Disagree with me if I'm wrong—I prefer accuracy over validation.
Prices in EUR, metric units, DD/MM/YYYY dates.
For code: include error handling, explain non-obvious parts.

Under 400 characters total. Covers the essentials without bloat.

The Actual Benefit

I used to spend 30 seconds per conversation setting context. That's 15 minutes a day if I'm using AI frequently.

Now I just ask my question. The AI already knows I want direct answers, European context, and technical depth without condescension.

That's not life-changing. But small friction removed adds up. The AI becomes a tool that actually fits how I work instead of a generic assistant I have to train from scratch every time.

Set it up once. Get better responses forever.