Inspired by the caveman skill that saves token counts within Claude by having it talk like a caveman, I thought perhaps having Claude leverage the vocabulary of baby sign language would save even more tokens.

I fed Claude a baby sign language dictionary and had it create a skill where it used that dictionary, allowing for technical words in backticks and code snippets too.

It unfortunately was not great, however it did come up with some exceptional phrases:

Technical Term	Babysign Translation
database host	big-book-home
max connections	most doors open same time
password	secret-word
JWT exp field	big-number-time (no small dots)
Date.now()	small-number-time (with small dots)
object reference	thing-with-names
re-render	make new picture
component (client)	baby home
idle timeout	close door after wait
min connections	keep doors open always
auth middleware	auth
unit mismatch	not same! like try put big apple in little cup
connection pool	pool for many want-things same time

Benchmark: Normal vs Caveman vs Babysign

Model: claude-sonnet-4-20250514
Trials: 1 per prompt per mode
Prompts: 3 of 10 (react-rerender, auth-middleware-fix, postgres-pool)

Token Counts

Task	Normal	Caveman	Saved	Babysign	Saved
Explain React re-render bug	1180	237	80%	265	78%
Fix auth middleware token expiry	704	125	82%	257	63%
Set up PostgreSQL connection pool	2272	635	72%	471	79%
Average	1385	332	78%	331	73%

Caveman: 72%--82% savings range.
Babysign: 63%--79% savings range.

Sample Responses

React Re-render Bug

Caveman (237 tokens):

Object reference change every render. React see new object, trigger re-render.
Common causes: Create object inline, return new object from function each time, spread operator create new reference. Fix: useMemo for computed objects.

Babysign (265 tokens):

Bad thing! You make new thing-with-names each time. Baby home think it other thing -> make new picture. Put thing outside or hold same thing with useMemo.

Auth Middleware Fix

Caveman (125 tokens):

JWT exp field in seconds. Date.now() in milliseconds. Fix: Date.now() < token.exp * 1000

Babysign (257 tokens):

Bad thing in auth! token.exp = big-number-time (no small dots). Date.now() = small-number-time (with small dots). Not same!
Like try put big apple in little cup. Fix: Date.now() < token.exp * 1000

PostgreSQL Pool Setup

Caveman (635 tokens):

PostgreSQL connection pool setup need pg package. max: 20 max connections, min: 5 min connections, idleTimeoutMillis: 30000 close idle after 30s...

Babysign (471 tokens):

Big-book need pool for many want-things same time. max: 20 most doors open same time, min: 2 keep doors open always, idle: 10000 close door after wait...

Observations

• Average tokens nearly identical: caveman 332, babysign 331
• Caveman wins on short prose answers (auth fix: 125 vs 257) -- dropping articles/filler is very efficient for terse explanations
• Babysign wins on code-heavy answers (postgres pool: 471 vs 635) -- vocabulary constraint forces harder prose cuts around
  unchanged code blocks
• Different compression strategies: caveman removes grammar; babysign constrains vocabulary to ~100 baby sign words, forcing
  creative paraphrase