Cole👽

AI agents, explained | Sourced, not hyped | Free guides

[Cole👽 @colerepo]

same 30 steps/model: $0.10 vs $0.37

four old files caused it. Claude Code resends context each turn; cached tokens count.

/usage, 1h vs 5m cache TTLs, /autocompact, and the 100K limit stopped my run.

10 steps + settings + read-only audit prompt ↓ https://t.co/0QUOATOBzx

[Cole👽 @colerepo]

@dotudev right, a line in CLAUDE.md is a request, a hook is enforcement. the agent can ignore a sentence, not exit 2

[Cole👽 @colerepo]

everyone's asking what a harness is this week.

here's the side you actually control. 7 parts, all files in your repo.

straight from Anthropic's guide for big codebases:

1. CLAUDE.md - the rules it reads every session
2. hooks - scripts that block bad moves before they run
3. skills - workflows you'd otherwise retype
4. plugins - hooks, skills, MCP and more, packaged into one install
5. MCP - your tools and services
6. LSP - real type errors and definitions, not grep guesses
7. subagents - side work that stays out of your main context

CLAUDE.md is just part 1.

the card shows all 7 as real files. send it to Claude Code and it drafts whatever your repo is missing.

the full course shows how to make it stop repeating the same mistake ↓

[Cole👽 @colerepo]

the engineers building Claude Code and Codex don't hit retry when an agent screws up.

they turn every failure into a rule, a hook, a skill or a test.

here's the whole system in 12 steps, with copy-paste hooks ↓ https://t.co/XcVsQzC83V

[Cole👽 @colerepo]

@TokenGremlin a security update isn’t a launch announcement tho

[Cole👽 @colerepo]

@XFreeze the real shift is from answering to actually doing

[Cole👽 @colerepo]

one "Say OK" after an hour away cost me 55.1k tokens

that's where your Claude Code limit goes when you step out. the cache dies after an hour and the whole chat gets re-read

2.1.304 finally warns you first

save these 2:

before a break:
/compact keep: <what matters>

on 2.1.305:
/usage → press f

where the rest of your limit goes, step by step ↓

[Cole👽 @colerepo]

same 30 steps/model: $0.10 vs $0.37

four old files caused it. Claude Code resends context each turn; cached tokens count.

/usage, 1h vs 5m cache TTLs, /autocompact, and the 100K limit stopped my run.

10 steps + settings + read-only audit prompt ↓ https://t.co/0QUOATOBzx

[Cole👽 @colerepo]

@pi_himself per turn is where it shows. by request 13 my run was re-reading 265k from cache every turn before doing anything new

[Cole👽 @colerepo]

same 30 steps/model: $0.10 vs $0.37

four old files caused it. Claude Code resends context each turn; cached tokens count.

/usage, 1h vs 5m cache TTLs, /autocompact, and the 100K limit stopped my run.

10 steps + settings + read-only audit prompt ↓ https://t.co/0QUOATOBzx

[Cole👽 @colerepo]

@pi_himself @BlueTearsPT exactly, that run was tiny 160-token steps. the 22k reads were the one that thrashed

[Cole👽 @colerepo]

@pi_himself do you track it per turn somewhere, or just watch /usage?

[Cole👽 @colerepo]

@harden_369741 swap those imports for .claude/rules files with a paths: line. that's when the number actually moves

[Cole👽 @colerepo]

everyone's telling you to delete your CLAUDE.md

don't. split it

imports won't save you. Anthropic's docs: they "don't reduce its context cost"

what does: rules that only load when Claude touches matching files

save as .claude/rules/api.md:

---
paths: "src/api/**"
---
- your API-only rules here

no paths: line = it loads every session anyway

whole CLAUDE.md at once? save the image → drop it into Claude Code → it sorts every rule into global / scoped / stale → moves nothing without your yes

CLAUDE.md is just one leak. here's where the rest of your limit goes ↓

[Cole👽 @colerepo]

same 30 steps/model: $0.10 vs $0.37

four old files caused it. Claude Code resends context each turn; cached tokens count.

/usage, 1h vs 5m cache TTLs, /autocompact, and the 100K limit stopped my run.

10 steps + settings + read-only audit prompt ↓ https://t.co/0QUOATOBzx

[Cole👽 @colerepo]

@harden_369741 lol imports are the diet soda of CLAUDE.md. feels lighter, same tokens

[Cole👽 @colerepo]

@BrethofAI nothing worse then Claude "fixing" the one weird thing that's weird on purpose

[Cole👽 @colerepo]

@pi_himself right, and in my run they were read before step 1, so all 30 steps carried them. thet's most of the 3.7x gap

[Cole👽 @colerepo]

@topxea curious how big yours got before you split it?

[Cole👽 @colerepo]

@pi_himself @BlueTearsPT on small steps it actually paid off, about 12.5% cheaper. with 22k file reads it refilled right after every compact

[Cole👽 @colerepo]

@TokenGremlin this is why the harness matters as much as the model

[Cole👽 @colerepo]

@Mr_Buzzoni curious how they measured that 10x improvement

[Cole👽 @colerepo]

@aksenHQ the ones that actually do stuff. commit, deloy, post. rest can stay auto

[Cole👽 @colerepo]

didn't know this until I checked /context: every skill you install rides along with every message you send

mine: 59 skills, 9.3K tokens

one line in a skill's frontmatter takes it out of context until you call it by /name:

disable-model-invocation: true

bonus: Claude can't fire that skill on its own anymore. Anthropic's docs say it straight: "You don't want Claude deciding to deploy because your code looks ready."

so start with skills that deploy, commit or send messages. then anything you only run by hand

one skill? paste this into Claude Code:

add disable-model-invocation: true to the frontmatter of my <skill-name> skill. change nothing else.

check: new session → /context → "Skills" drops

too many to do one by one? save this image and drop it into Claude Code

it finds every skill, counts what each one costs you and asks before editing anything

and that's only skills. here's where the rest of your Claude Code limit goes ↓

[Cole👽 @colerepo]

same 30 steps/model: $0.10 vs $0.37

four old files caused it. Claude Code resends context each turn; cached tokens count.

/usage, 1h vs 5m cache TTLs, /autocompact, and the 100K limit stopped my run.

10 steps + settings + read-only audit prompt ↓ https://t.co/0QUOATOBzx

[Cole👽 @colerepo]

@pi_himself @BlueTearsPT yep, saw exactly that 100k. It compacted 3 times in a row and then just stopped to run