- If you want to know why your AI gets worse when you add more rules, the short answer is that stacked instructions start to conflict, repeat, and cancel each other out.
- Every bad output tempts you to add one more rule. That patch often makes the next answer worse, not better.
- You do not need a bigger setup. You need a short audit.
- Ask five questions of every rule: is it already default, does it conflict, does it repeat, is it a one-off patch, and is it too vague?
- Keep the setup simple as your business grows, and stay the person who decides what stays in.
You give the AI a clear instruction. It misses. You add a rule so it will not miss again. It misses in a new way, so you add another rule. A few weeks later the setup is long, the answers are worse, and it feels like the tool is getting dumber.
That is why your AI gets worse when you add more rules. It is not always the model. It is often the pile of instructions sitting on top of it.
I keep my own AI setup simple on purpose. The job of a rule is to help the next answer, not to patch the last one.
Why does my AI get worse the more instructions I give it?
Most people treat custom instructions like a punch list. Something goes wrong, they add a line, and they never take a line out.
The model has to satisfy all of those lines at once. When two of them pull in different directions, it has to guess which one you care about. When several of them say the same thing in different words, it burns attention repeating itself. When a line is vague, it fills in a meaning you did not intend.
That is how a well-meant setup turns into noise. You were trying to get more control. You accidentally made the request harder to follow.
If you want a clean brief for a single task, I walk through that in how to give AI clear instructions. This post is about the standing rules that sit behind every chat, not the one prompt you write today.
What happens when AI rules conflict or repeat each other?
Conflict is the quiet killer. One rule says "keep it short." Another says "give full context and examples." One says "be direct." Another says "be warm and thorough." The model will try to do both, and the result often feels mushy, or it drops one of the rules without telling you.
Repeats are almost as messy. "Write like a founder," "sound like me," and "stay in my voice" may all mean the same thing to you. To the model, that is three instructions to juggle.
You do not need a paper to see this. You can feel it when an answer ignores a rule you were sure you had covered. The rule is probably still in there. It just lost the fight with a newer one.
How many custom instructions is too many?
There is no magic number I can give you, and I will not invent one. Too many is when the setup has rules you cannot remember, rules that fight each other, or rules you added after a single bad answer and never reviewed.
If you cannot read your custom instructions out loud in one short pass and know what each line is for, it is too long.
A short, clear setup will beat a long, careful one that no longer fits. If you need a refresher on the basic ChatGPT setup, I cover the starting point in ChatGPT custom instructions. Use that as a base, then cut.
What are the 5 questions to audit your AI setup?
Open your custom instructions, your project rules, or your agent brief. Walk every line through these five questions. If a line fails one of them, cut it or rewrite it.
- Already default? Is this already how the tool behaves? If you are telling it to be helpful, clear, or accurate, you are using space for something it already tries to do.
- Conflicts? Does this line pull against another line? If yes, keep the one you mean and delete the other. Do not leave both and hope.
- Repeats? Are you saying the same thing twice in different words? Keep the strongest version.
- One-off patch? Did you add this because of one bad output? If that situation is rare, take the line out and handle it in the prompt when it comes up.
- Too vague? Could two people read this and do two different things? "Be more professional" is vague. "Write in short sentences, no jargon, and end with one next step" is usable.
When you finish, you should have fewer lines than you started with, and each one should earn its place.
How do I keep my AI setup simple as my business grows?
Your business will change. Your offers, your voice, and your tools will change too. If you only add rules and never remove them, the setup will always lag behind the work.
Set a simple habit. When you add a rule, look for one to remove. When you start a new project or agent, do not copy the entire pile forward. Start from the short version, then add only what that job needs.
If you are setting up an agent, not just a chat, start with standards and one job, not a giant rulebook. I walk through that in how to set up an AI agent for your business.
And keep safety separate from style. Limits that stop an agent from sending, deleting, or spending without you are not the same as stacking tone rules after every missed sentence. You want the first kind. You want less of the second. I break down those limits in AI agent guardrails.
How do you stay in charge of the setup instead of patching it?
This is the human-first part. You own the setup. The tool does not.
A bad output is feedback, not a reason to tack on another standing rule. Read the answer, decide what was wrong, and fix the brief for that task. Only promote a line into your standing rules if you will need it often, and only after it has been through the five questions.
You stay in charge when you can explain every rule in your own words. If you cannot, the setup is running you.
That is the whole fix. Audit the pile, cut what does not earn its place, and add new rules slowly. Your AI gets worse when you add more rules that fight each other. It gets better when the rules you keep are few, clear, and yours.
Frequently Asked Questions
Should I delete all of my custom instructions and start over?
If you cannot tell what each line is for, yes, start from a short setup and add back only what you still need. If you can already name the purpose of every line, run the five-question audit first and keep what passes.
Will a shorter setup make my answers worse?
A shorter setup makes answers worse only if you cut a rule the model actually uses and you still need. Most of the time, cutting conflicts and patches makes the answers cleaner.
Is this the same as adding guardrails to an AI agent?
No. Guardrails are limits on what an agent is allowed to do, like sending a message or changing a file. This audit is about the standing instructions that shape how the AI writes and thinks. Keep the limits. Cut the clutter.
How often should I audit my AI instructions?
Any time the answers start to drift, and any time you add a new tool, project, or offer. A short pass when something feels off is enough. You do not need a big quarterly project.