most of your AI's budget goes to re-reading its own paperwork. every single message.

we fixed that.

65% of your config is waste
215K tokens you get back
0 accuracy loss

every message you send, these three things load first

28K
system prompt
fixed. you can't touch it.
2–18K
CLAUDE.md
yours. this is where the waste lives.
1–8K
MEMORY.md
grows forever. nobody prunes it.
then your actual work gets whatever's left.
4,500
tokens per message on config alone
×
50
turns in a session
=
225,000
tokens that never change, re-sent every time

drop your CLAUDE.md

drop your CLAUDE.md here
or click to browse
paste instead try with sample
why it works

your AI doesn't read. it counts.

grammar, formatting, politeness — all tokenized, all billed, none of it necessary.

before 2,847 bytes
## Execution Model
- Use Opus for all decisions,
  planning, and code generation
- Use Sonnet only for fetching
  data, reading files, and research
- Never let Sonnet make edits or
  architectural decisions
- When stuck, fork a parallel agent
  with an alternate approach instead
  of retrying the same thing

## Agent Behavior
- Skip preamble, go straight to
  action on every response
- Complete implementations, never
  leave stubs or TODOs
- Test every build before reporting
  success to the user
- Flag context degradation instead
  of silently getting worse
→
after 418 bytes
exec:opus decides+writes+edits
  sonnet fetches only—zero edits
  stuck→fork parallel,don't loop

conv:skip preamble→action
  complete impls not stubs
  test builds|flag context rot
10/10 accuracy retained
what we actually do

six things. by hand. once.

no tool does this. every config is different.

01
config compression
rewrite CLAUDE.md in compressed notation. strip grammar, merge redundant sections, eliminate descriptions the model already knows.
18,129 B → 3,172 B — same comprehension, 82% smaller
02
memory architecture
restructure memory index and files. remove stale entries, merge overlapping records, tighten descriptions so the right ones actually load.
40% of memory entries are unused on average
03
deduplication
find and eliminate repeated instructions across config, memory, and system prompt. one source of truth per instruction.
most setups have 2-3x duplicate instructions
04
execution model
define when your AI should act vs ask, which model handles what, how to parallelize agents. the behavioral contract.
opus thinks. sonnet fetches. agents fork. loops die.
05
skill pruning
audit installed skills and MCP servers. remove unused, compress descriptions of what you keep. each skill description burns tokens every message.
skill descriptions alone can burn 10K+ tokens per turn
06
context budget
set guardrails for session length, agent spawning, and context rotation. teach your AI to flag when it's degrading instead of silently getting worse.
context rot = why long sessions degrade silently

questions

does this actually work or is it a gimmick?
we ran A/B tests. same agent, same questions, verbose config vs compressed. 10/10 both times. every time. the AI doesn't care about your formatting — it cares about tokens. grammar is decoration. strip it.
will it break anything?
no. same instructions, fewer tokens. if you compress too aggressively and comprehension drops, you'll know immediately — just back off that section. it's a text file, not surgery.
only for Claude Code?
nah. works anywhere there's a system prompt or config — Cursor, Windsurf, Aider, whatever. if tokens cost money or count toward a limit, this helps.
how long does this take?
the tool above takes one second. doing it properly by hand, maybe 30-60 minutes for a typical CLAUDE.md. memory files take longer. the page above shows you exactly what to cut.
my limits still hit though?
eventually, yeah. but way later. 65% less overhead means 65% more runway before your session turns to mush. and if you set up context rot detection, your AI tells you when it's degrading instead of quietly getting dumber.
who made this?
marow.ai — not a startup. not raising. just building things that work.

that's it. go use it.

everything here is free. no signup, no newsletter, no "join our community." just smaller configs and longer sessions.

marow.ai also bubbleshop.ai