Why I ran this experiment
Give a bot the same instruction and some days the result is spot on, other days it looks like only half was read. I suspected the problem was not the wording but where the instructions lived, so I split them across three places for a week. I ran everything on a Windows device and a Mac mini, 5 runs per location, 15 in total.
The three locations
• **A. Chat window**: paste the instructions every time
• **B. Notes app**: write them in a note and send only the link
• **C. Work folder**: save them as a CLAUDE.md file inside the folder
Results
• A: the longer the conversation ran, the fuzzier the early rules got. In 2 of 5 runs, a rule or two went missing.
• B: the bot sometimes could not open the link and guessed the content instead. That happened in 3 of 5 runs.
• C: the file was read automatically at the start of every session, and all 5 runs followed the rules.
Telling the bot works less well than putting the instructions where it looks the moment it wakes up.
Try it yourself (5 minutes)
1. Create a folder for your work.
2. Add a new CLAUDE.md file inside it.
3. Write five lines of rules or fewer.
4. Start Claude Code in that folder and ask, 「What rules are active right now?」 to confirm.
# Rules
• Always answer in Korean
• Tell me before editing any file
• When done, list the names of the files you changed
Longer rule lists actually led to more misses, so keeping only what truly matters worked better.