if you use more than 1 AI, you probably copied the same instructions file into all of them
i did too. then i had each of my AI CLIs look up its own official docs and tell me what it needs from me
they don't all read the same file, format or folder. and they don't all read the “whole thing” 👀
why one big rules file breaks
every AI tool has its own rules about instructions:
→ which file it looks for
→ where that file has to live
→ how much of it it will load
the last one is the sneaky one. Antigravity (Google's AI coding CLI that runs Gemini) reported from its own docs that it stops reading a rule file past 24,000 bytes.
Grok, on the other hand, reported no size cap at all. same instructions, totally different limits
my main rules file was 88,682 bytes. copied over as-is. when I started looking into it yesterday, I realized why sometimes it feels like it ignores random instruction it knows to follow!
[have you ever checked how much of your claude[.]md or agents[.]md your AI actually loads or reads? ask your AI]
have each AI read its own manual
so i sent 5 of my AI CLIs the same job at the same time:
read your own official, current and up-to-date release docs, API docs and change logs then tell me why you skip or ignore instructions sometimes. also report back on what kind of instruction file you load, where it lives, and quote the sourceeach one came back knowing a limit the others didn't. they even caught each other's mistakes (one flagged that another AI's instruction file pointed at the wrong folder)
then I told them to optimize their own instruction files. keeping my personal context they need to know intact and truncating instructions to make them more ‘AI-friendly’
Codex went from 88,682 bytes to 26,938 bytes, and in a fresh chat it still quoted my standing rules word for word
you can run the same thing. have your agent do this:
read your own official, current docs (release notes, change log, API docs, official blogs etc.) tell me which instruction file you load, where it has to live, how much of it you load, and what you need from me in it. quote the source for each answer
run a test
this part works even if you only use 1 AI
open a fresh chat with no tools and ask:
quote the 3 most important rules from your instructions file, word for word
if it quotes them, the file loads. if it can't, or it makes rules up, that file probably isn't reaching it
why no tools? with tools it can go open the file and read it to you, which proves nothing about what it carries into every chat
my favorite answer came from Hermes, the one AI i ran with no tools at all. instead of guessing, it wrote "UNVERIFIED" for every section. most trustworthy report because that means it didn’t hallucinate LOL
what to do today
ask your AI which instruction file it reads and how much of it loads every session
run the 10-second test in a fresh chat
if anything comes back cut off or made up, shorten that file first
your rules only work if they load. run the test before you add another one ✌🏽
I’ve been putting a lot of my processes and workflows up as GitHub repos and skills on Agensi. The repos are free and easy npm installs.
Feel free to check em out and hope you find em useful 🫶🏽
want me to set this up for you? book a call at aunysillyme.com and we'll get every AI you use loading what it should.
SEE YA ON THE NEXT ONE ♡