Caveman mode
A hook that compresses the agent's replies by about 75% — same technical substance, none of the ceremony.
What it is
Caveman mode is a style filter on the agent's output. It strips articles (a, an, the), filler words (just, really, basically, actually, simply), pleasantries (sure, certainly, of course, happy to help), and all hedging — then speaks in tight fragments using short synonyms: big not extensive, fix not implement a solution for.
Every technical term, code block, and error message is kept exactly as-is. Roughly three-quarters fewer output tokens for the same information. The model doesn't get dumber — it just stops padding.
Same signal, less noise
The before/after is stark. The words that vanish are the ones that were never load-bearing.
Sure! I'd be happy to help. The issue
you're experiencing is likely caused by
the auth middleware
— specifically, the token expiry check
which is currently using
< instead of <=.
Let me go ahead and fix
that for you:
Bug in auth middleware. Token expiry check use < not <=. Fix:
Drop the fluff, keep the facts — and the code, verbatim.
The rules
- 1
DROP: articles, filler, pleasantries, hedging
Fragments are fine. Short synonyms preferred:
bignot extensive,fixnot implement a solution for. Pattern:[thing] [action] [reason]. [next step]. - 2
KEEP: technical terms, code blocks, error text — exact
Nothing inside a code fence is touched. Error messages are quoted verbatim. Technical vocabulary is never paraphrased down to something vague.
- 3
Intensity:
/caveman lite|full|ultra— "stop caveman" to revertlitedrops only obvious filler;full(default) is aggressive;ultrapushes to near-single-word fragments and includes wenyan variants. Toggle any time mid-session. - 4
Auto-clarity: it switches off when ambiguity is dangerous
Security warnings, irreversible-action confirmations, and multi-step sequences where fragment order could be misread are written in plain English automatically. Caveman resumes after. Code, commit messages, and PRs are always written normally.
Why bother
Output tokens cost real money and real time. Most of an agent's prose is ceremony, not information — the pleasantries, the hedges, the wordy constructions that say the same thing in three times as many words. Compress the prose; keep the signal. At high usage, a 75% reduction in output tokens is not a rounding error.
The guardrail matters too. Anything where ambiguity is genuinely dangerous — a security warning, a destructive-operation confirm, a numbered sequence that must be followed in order — is written in plain, complete sentences. And anything that ends up in the repository — code, commit messages, pull requests — is always written normally. Caveman mode is a response style, not a code style.
How I actually use it
My take
I read fast and pay per token, so full is my default. I keep it
on almost all the time — the only things I actively switch back for are the moments caveman
already handles automatically (security stuff, irreversible ops). Everything else can be a fragment.
Honestly, once you've worked with it for a day, the normal chatty agent voice starts to feel
like someone clearing their throat before every sentence. This entire Agentic coding section,
in fact, was built with the agent in caveman mode.