Sebastian Schkudlara
There is a mode in Lope called intelligent caveman. It is on by default. It cuts the token cost of every validator response by 50 to 65 percent without losing a single line number, path, or error message. It is also a very stupid joke, and I’m going to explain why the stupid joke matters.
The wordplay
“Intelligent caveman” is a wink. The mode sounds primitive — drop articles, grunt in fragments, say as little as possible. The compression is actually linguistically precise. Code, paths, line numbers, error messages, commands, and hash digests stay exact. Only the wrapper prose gets grunted away. Terse like a caveman, smart like a linguist.
The whole trick is a single directive injected at the top of every validator prompt:
Respond terse. Drop articles (a, an, the), filler (just, really, basically), hedging (I think, perhaps, might), pleasantries (sure, happy to help). Fragments OK. Keep exact: code, paths, line numbers, commands, error messages. Rationale = 1 to 3 short sentences, not paragraphs.
That’s it. No fine-tuning. No post-processing pipeline. Just a prompt prefix that tells the validator to stop writing Harvard undergraduate essays when a senior engineer would grunt and point at the line.
Before and after
Without caveman mode (LOPE_CAVEMAN=off):
rationale: I've reviewed the proposal carefully and I think there are a few
issues that should probably be addressed before moving forward. Specifically,
I noticed that the authentication middleware doesn't appear to handle the
case where the token has just expired, which could potentially cause some
edge case failures in production. Additionally, I'd suggest adding some
rate limiting to prevent abuse of the refresh endpoint.
Roughly 60 tokens.
With caveman mode (LOPE_CAVEMAN=full, the default):
rationale: Auth middleware misses expired-token edge case. Add rate limit on
/refresh endpoint. Fix at middleware/auth.go:142.
Roughly 22 tokens. Same information. 63 percent fewer tokens.
Notice what stayed exact: middleware/auth.go:142. The file path and the line number did not get “grunted.” The compression is deliberately asymmetric. Human prose is compressible. Technical identifiers are not.
Why this matters at scale
Lope fires N validators × M phases × up to 3 retries per sprint.
A typical sprint has 3 validators, 4 phases, and an average of 1 retry somewhere along the way. That’s 36 validator calls. Multiply 60 tokens by 36 = 2,160 tokens of pure wrapper prose. Multiply 22 tokens by 36 = 792. 1,368 tokens saved per sprint, per user.
On a local validator (Ollama, llama.cpp) the savings are latency. On a paid API the savings are money. It compounds fast at team scale.
Three modes
| Mode | Env var | When to use |
|---|---|---|
full (default) |
LOPE_CAVEMAN=full |
Maximum compression. Drops articles, filler, hedging. Fragments OK. |
lite |
LOPE_CAVEMAN=lite |
Drops filler and hedging only. Keeps full sentences. |
off |
LOPE_CAVEMAN=off |
Disable entirely. Use when debugging lope itself. |
LOPE_CAVEMAN=lite lope negotiate "build auth"
When to turn it off
Rarely. Only if the validator’s prose itself is the deliverable — say, reviewing a board memo where you want polished sentences. For everything else (code review, sprint validation, budget checks, methodology review) keep it on.
Credit
Core terseness rules from JuliusBrussee/caveman (MIT). Lope’s contribution is integrating those rules into the validator prompt injection pipeline where savings compound across multi-round ensemble review.
The joke is load-bearing. Terse prompt, exact payload, measurable cost savings. Grunt at the AI. Count the tokens you didn’t pay for.
Read https://raw.githubusercontent.com/traylinx/lope/main/INSTALL.md and follow the instructions to install lope on this machine natively.
— Sebastian