AI basics · Part 1
Why do I get back something I could have written myself?
How to stop AI giving you generic answers
It is not the model. It has never been told anything about you.
AI usually gives generic answers when it knows nothing about you, your organisation or the job you are doing. Add clear profile instructions once, then add project-specific context where it belongs. You will spend less time repeating yourself and get answers that fit your work.
Why AI answers are generic
AI learned from an enormous amount of writing. Almost none of it was about your organisation.
It does not automatically know what you sell, who your customers are, how you make decisions or how you sound. You know all of that. You can teach it the parts that matter.
A capable model with no useful context still has to guess.
What instructions do
An instruction is information or guidance that applies automatically, without you retyping it in every chat.
The tool reads those instructions alongside your new request. That context shapes the answer that follows.
Claude calls these profile instructions and project instructions. ChatGPT calls the account-level version Custom Instructions. Gemini calls them Instructions for Gemini under Personal Intelligence.
How to write a useful instruction
Write a useful instruction by explaining, in one or two sentences, what you do, who you help and what matters in your work.
Imagine you have just met someone new. In 20 seconds, how would you tell them that? Write it down in plain language. Start short. Add detail only when the answers show you something is missing.
I run a medical practice in Chatswood with 4 GPs and 3 admin staff.
Where instructions live
Instructions live at organisation, account and project level. Put each one at the narrowest level where it should apply.
-
Your organisation
Claude organisation settings > Organisation instructions
Never give clinical advice.
-
Your account
Claude Settings > Profile > Instructions for Claude
I run a medical practice in Chatswood with 4 GPs and 3 admin staff.
-
One project
Inside a Claude project > Set project instructions
We are digitising the onboarding of new patients.
In ChatGPT, account-level instructions live under Settings > Personalization > Custom Instructions. In Gemini, they live under Settings & help > Personal Intelligence > Instructions for Gemini. Microsoft Copilot has an equivalent, though Microsoft still labels it a preview feature, so expect the menu to move. Organisation controls depend on your plan.
How instructions shape one answer
When you send a message, the model reads organisation instructions, profile instructions, project instructions and your new request together.
Here are examples of the current onboarding process. Help me refine it.
What the model reads
- Never give clinical advice.
- Organisation instructions
- I run a medical practice in Chatswood with 4 GPs and 3 admin staff.
- Profile instructions
- We are digitising the onboarding of new patients.
- Project instructions
- Here are examples of the current onboarding process. Help me refine it.
- What you just typed
Prompting advice keeps changing
Old prompting tricks often make answers worse now. Give the model clear context, a clear task and a useful example instead.
When these tools first landed, prompts were often long, elaborate and full of tricks. Models have improved. If your prompting habits are a year old, test whether they are still helping.
If your prompting habits are a year old, test whether they are still helping.
Try this in five minutes
Compare the same question with and without profile instructions to see how much context changes the answer.
-
Open a new chat and ask a broad question.
How can I grow my business?
-
Read the answer and notice where it had to guess.
-
Add a short profile instruction describing your organisation and role.
-
Open another new chat and ask the same question.
-
Compare the two answers. Keep only the instruction that made the second one more useful.