Claude is not following your instructions because, to Claude, your instructions are advice. They are not settings. Anthropic says so in its own documentation: Claude treats your rules "as context, not enforced configuration."
There are a lot of ways to get this wrong, and I have tried most of them. I wrote longer rules, then louder ones, then the same rule in three places. The repeat corrections kept coming.
A quick word on names, because they confuse everyone. Your rules are the standing instructions Claude reads before every chat. In the Claude app and Cowork, they are called project instructions. In Claude Code, they live in a file called CLAUDE.md. I will just call them your rules.
Here is what fixes it, whether you use the Claude app, Cowork or Claude Code, so you can skip those steps. Most rules should stay as short, plain guidance. A few should be checked by a small piece of code that runs every time. And any correction you make twice is a sign the rule is in the wrong place.
Using the Claude app or Cowork? Start here
Cowork is the mode in the Claude desktop app that works on tasks and files for you. Your rules live in project instructions. Keep them short, specific and at the top. Here is a block you can paste and edit:
Rules for this project. 1. Write in plain words a busy person can read fast. 2. Never use em dashes. 3. Every number needs a source link, or write [INPUT NEEDED]. 4. Ask before sending anything to another person. 5. Put the answer first, then the detail.
A few clear rules work better than many vague ones. Anthropic's help page describes project instructions as a way to "help guide how Claude works" (Anthropic, Cowork help), the same idea as the rest of this piece: guidance, not settings. If even your short rules get missed, jump to "What if Claude still misses my rules in the app or Cowork?" below.
Why is Claude not following the instructions in my CLAUDE.md file?
A CLAUDE.md file is a plain text file of standing instructions for Claude Code, the version of Claude that works directly with files on your computer. Claude reads it at the start of every session.
The catch is how it reads it. Anthropic explains that the file arrives as a message after Claude's built-in instructions, and that "there's no guarantee of strict compliance, especially for vague or conflicting instructions." (Anthropic, memory docs)
So your rules land the way a note from you lands. Claude weighs them alongside everything else in the conversation. Anthropic's best practices page is honest about the result: a rule like "always typecheck before you finish" in this file "works most of the time, not every time." (Anthropic, best practices)
Most of the time is fine for a style preference. It is not fine for "never send a client anything without my approval."
Does a long list of rules make Claude worse at following them?
Yes. Anthropic's guidance is to keep each CLAUDE.md file under 200 lines, because longer files use up more of Claude's working memory and can reduce how closely it follows them. (Anthropic, memory docs)
My main CLAUDE.md file is 266 lines. With the two project files that load beside it, Claude starts every session with 5,390 words of my rules before it reads my request. Every line is there because something went wrong once. Together, they make each line easier to miss.
Outside research points the same way, with limits worth knowing. A 2023 study found AI models pay the most attention to the start and end of a long input, and the least to the middle. (Liu et al.) A 2025 test gave 20 AI models up to 500 instructions at once, and the best one followed 68% of them. (IFScale) Neither tested real business rules, so treat them as a direction, not a number.
Two smaller traps hide in long files. When two rules disagree, Anthropic warns that Claude "may pick one arbitrarily." And vague rules fail more often. "Use 2-space indentation" works. "Format code properly" does not.
Why does Claude follow a rule for a while and then stop?
There are three usual causes.
Long conversations get compressed. When a session nears its memory limit, Claude Code summarizes the earlier conversation to make room. Anthropic says the main CLAUDE.md file in your project folder is read again after that happens, so instructions you only typed into the chat are the ones at risk. The fix is to put lasting instructions in the file. (Anthropic, memory docs)
Skills load only when they are used. A skill is a saved set of instructions for one kind of task, like "write a blog post for my brand." Only its short description loads up front, and "full skill content only loads when invoked." (Anthropic, skills docs) A rule inside a skill does nothing until that skill runs.
A rule can be mentioned without being required. This one caught me. My content review skill referred to "the panel review" by experts that every draft should get. No step actually said to run it. So the panel ran when a session happened to remember, and not otherwise.
Does writing IMPORTANT in capital letters help Claude follow a rule?
Usually not. Anthropic's best practices page says "adding IMPORTANT probably won't fix it," and that with current models "piling on emphasis mostly makes Claude over-apply the rule." (Anthropic, best practices)
So shouting can backfire. Claude may start applying the rule in places you never meant.
The same page gives the best test I have found: if you keep adding emphasis to a rule and Claude still misses it, "that's a sign the line belongs in a hook."
What if Claude still misses my rules in the app or Cowork?
The app and Cowork do not give you the hooks described below, so the fix is to make Claude check its own work against your rules before you see it. Four moves that help:
- Ask for a self-check. End important requests with: "Before you show me, check your answer against my project rules and fix anything that breaks them. List any rule you could not meet."
- Repeat the one rule that matters most in the request itself, not only in the project instructions.
- Save your best request as a reusable prompt with the rules built in, so you are not retyping them.
- Check the one or two things that would hurt most if they were wrong, like a number or a recipient, before anything leaves your hands.
This still is not a guarantee. If a rule truly can never break, that is the point where a few hooks in Claude Code earn their keep.
App and Cowork users: the next three sections are about Claude Code. Skip to "Do this today" if that is not you.
What is a Claude Code hook, in plain English?
A hook is a small script that Claude Code runs on its own at a set moment. Anthropic describes the point as "deterministic control: certain actions always happen rather than relying on the LLM to choose to run them." (Anthropic, hooks guide) In plain words, the computer does the checking, so nobody has to remember.
The moments I use most:
- Before Claude saves a file or runs a command. The hook can stop it.
- When I send a message. The hook can add a reminder before Claude replies.
- When Claude tries to finish. The hook can send it back to work.
To stop an action, the script sends back a stop signal (Anthropic calls it exit code 2). One thing to know: if the script itself breaks, it does not stop anything, so test each hook once. (Anthropic, hooks reference)
You do not need to write the script. Describe the rule in plain words and ask Claude to build the hook. That is how all of mine were made.
What changed when I moved rules into hooks?
Here is the clearest example, and it is a boring one.
I have a rule that files holding passwords and API keys never go into a folder that syncs to the cloud. It was written in my rules, and Claude broke it 14 times. Another of my rules told Claude where to put new projects, and that folder syncs to the cloud. Two of my own rules disagreed, and the wrong one kept winning.
Now a hook checks every file save and copy command before it runs. The argument is over.
The second example is why I take this seriously. A batch of 24 approved LinkedIn replies once went out from a real person's account in 67 seconds through a quick script. One message that had been pulled from the batch went anyway, and a note meant only for me went out with it. Now a hook blocks any LinkedIn send that skips the reviewed, spaced-out path. A mistake you cannot take back should never depend on memory.
My hooks now block em dashes in my writing, passwords in cloud folders, rushed LinkedIn sends, Word files left where I want Google Docs, and content drafts that skipped the expert panel. Each one started life as a written rule that kept getting missed.
Which rules belong in a hook and which belong in my written rules?
Keep the hook list short. Anthropic notes that if you find yourself with dozens of enforced rules, "most are probably preferences that Claude could have weighed on its own." (Anthropic, best practices)
A rule can live in one of three places:
- Your written rules. For taste and judgment: voice, formatting, how options are presented. Keep them short and specific.
- A step inside a skill. For process: "every draft gets four expert reviews." Have the step leave proof, like a saved review file.
- A hook. For the few things that must never break and that a computer can check: a file type, a word, a folder, a send. A hook can also check for the proof a skill step left behind, which is how a process step becomes truly required.
To decide, ask three questions. Have you corrected it twice? Would breaking it cost money, privacy or trust? Can a computer check it with a yes or no? Two yeses means it needs more than a written line.
This might not be for you if you are new to Claude. Write your rules first, live with them for a few weeks, and let the repeat mistakes tell you which few rules deserve a hook.
How do I stop correcting Claude for the same thing over and over?
Treat every repeat correction as a bug report about your setup. This is the loop I run:
- Find the rule you thought you had. Ask Claude to search your rules and quote it back.
- Name why it was missed. Usually one of five things: the file is too long, the rule is vague, two rules disagree, it was only said in chat, or it was never written down.
- Move it to the right place. Rewrite it short, make it a skill step with proof, or turn it into a hook.
- Log it. Keep a one-line-per-correction file and look for repeats once a week. Skip this if you only correct Claude now and then.
Step one surprised me. I was sure I had told Claude to make every document a Google Doc. Here is what I had actually said. Early on, I asked for documents I could open with one click, not plain text (.md) files, and I asked Claude to add that to my rules. I also said everything should live in Google Drive. I never actually said the words "make every document a Google Doc." A Word file opens with a click and sits fine in Drive, so it met both of my requests. Documents started showing up as Word files, and each time I corrected Claude in the chat. But a correction in the chat only covered that one document. When I went looking, I found that when Claude wrote my rule down, it wrote "Word files."
So Claude was following what was written down, not what I meant. A lot of the time, that is the whole problem. Be specific, and when Claude keeps missing something, check your rules first.
Do this today
If you use Claude Code: open your CLAUDE.md and count the lines. Over 200? Move topic rules into skills and cut one-time preferences. This might not suit you if you rely on a big setup, so start by moving rules rather than deleting them. Then pick the correction you have made most often and ask Claude to turn it into a hook.
If you use the Claude app or Cowork: cut your project instructions down to the few rules that matter most, put them at the top, and end your important requests with the self-check line above.
Either way, you get back the time you spend saying the same thing twice. I wish someone had told me sooner.
I write a Field Notes newsletter on running real work with AI, and you can subscribe here. If you want help setting this up for your team, I offer a free 45-minute AI strategy call, and you can book it here.
Which rule do you find yourself correcting most often?
Frequently asked questions
Why is Claude not following instructions?
Claude is not following instructions because it treats them as guidance, not settings. Anthropic's documentation says your rules are "context, not enforced configuration," with no guarantee of strict compliance. Long rule files, vague wording, rules that disagree, and instructions given only in chat all raise the miss rate. For rules that must never break, use a hook.
Why is Claude Code not following my CLAUDE.md file?
Claude Code most often misses CLAUDE.md rules when the file is long, vague or contradictory. Anthropic recommends under 200 lines per file and warns that when two rules conflict, Claude "may pick one arbitrarily." Rewrite rules to be short and checkable, remove conflicts, and turn any rule that keeps getting missed into a hook.
What are Claude Code hooks?
Claude Code hooks are small scripts that run automatically at set moments, such as before Claude saves a file or when it tries to finish. Anthropic says they give "deterministic control," so certain actions always happen instead of depending on Claude to remember. You can ask Claude to write a hook for you in plain words.
Why is Claude ignoring my skills?
Claude may ignore a skill because only its short description loads at the start of a session. Anthropic says the "full skill content only loads when invoked," and the description is how Claude decides when to use it. Make the description name the exact situations the skill is for, and have key steps leave proof a hook can check.
Does Claude remember my instructions after a conversation is compacted?
Claude Code re-reads the main CLAUDE.md file in your project folder after it compresses a long conversation, according to Anthropic. Instructions you only typed into the chat are the ones at risk of being lost. If an instruction should last beyond the current task, add it to your CLAUDE.md file or, for a rule that must never break, a hook.

