There is a specific kind of annoyance that only happens with AI tools.
You write a rule. A clear one. You tell the agent exactly how to handle the thing it keeps getting wrong. You save the file and feel like you fixed it.
Then you watch it do the wrong thing anyway. Not once. For weeks. Cheerfully, every time, while your perfectly good instruction sits three files away doing nothing.
If you keep any kind of instruction file for an AI, a CLAUDE.md, a system prompt, a page of house rules, you have felt this. The model nods at your rule. Then it does whatever it wants.
What took me too long to see is that none of this is random.
Some of my rules fire every single time. Others get quietly skipped no matter how clearly I phrase them. For a long stretch I could not tell which was which until something broke and I went digging.
So I started tracking which rules actually change the output and which are decorative. After enough of them, a pattern showed up. Broken instructions fail in five distinct shapes.
Here is the opinion that took me the longest to accept. Blame the model and you will be wrong almost every time. Look at the instruction instead. A rule that reads like a motivational poster gets treated like one.
1. The aspiration
Biggest category, and the hardest to give up, because these rules feel the most important.
Be thorough. Write clean code. Always think carefully. Prioritize quality.
They sound like principles. They are wishes. Nothing in them gives a model something to measure against, so it agrees with every one and changes nothing. A mood in the shape of a rule is still a mood.
2. The buried rule
Sometimes a rule is genuinely good and still never fires, because it is drowning.
Once your instruction file grows to hundreds of lines, every rule competes with every other rule for the model's attention. Whatever matters right now sits in a wall of things that do not. Correct and invisible turn out to be close cousins.
A good rule in a bloated file is a good rule you will not benefit from.
3. The positive-only rule
This one surprised me. Telling the model what to do is weaker than telling it what not to do.
Improve the code reads as open season. Do not rename anything I did not ask you to rename is a fence. A model walks around a suggestion easily. Walking through a wall is much harder.
Negative constraints bind harder than positive guidance, and most people write almost none of them.
4. The stale rule
Quietest failure of all. Your rule was true the day you wrote it and sits silently false now.
It names a file that moved. A flag that got renamed. A step that no longer exists. Your agent follows it faithfully, straight into the wrong place, and hands you a confident wrong answer that traces back to your own instruction.
A stale rule does more damage than no rule, because you trust it and it lies.
5. The unenforceable wish
Last category, and the one I fought hardest.
Always run the tests before you say it is done. Never commit secrets. Always update the docs.
Every one of these is correct. Every one asks the model to remember something on its own, with nothing external to trigger it. That holds until the single run where it matters, which is the only run you were counting on it.
Rewording it will not help. Wiring it will. When a rule has to hold every time, it stops being a rule and becomes a check that runs whether anyone remembers it or not.
What the working rules had in common
Once I sorted my rules this way, the split jumped out.
Working ones read like physics. Each carried a trigger, a specific behavior, and something that would notice if it broke. Failing ones read like aspiration. Good heart, no mechanism.
My rule library got big enough that I stopped being able to recite it from memory. The only ones I trust anymore do not depend on me, or the model, remembering to care.
Here is the uncomfortable part. Rules people are proudest of are usually aspirations. Clean-code, be-thoughtful, always-consider. They feel like leadership and do almost nothing.
Boring, negative, mechanically-checked rules are the ones shaping the output. Writing better rules is the wrong goal. Ask which of your current rules would survive being tested.
Your turn
Which of your AI rules do you secretly suspect is decorative?
If this was useful
I work through this in public, the wins and the freezes both, mostly on LinkedIn and YouTube. If the real version of building in the open is useful to you, that is where it lives. Find me on X, GitHub, and the work at next8n.com.
Top comments (0)