As we move from free/cheap AI Tokens to a more costly model I find myself looking for ways to use fewer tokens but still have the quality products I am looking for. I was turned on to a really creative solution to the problem called “Caveman Mode”.
Tokens are consumed parsing the data you send in and generated all the output data. That includes the details and instructions that come back with your code requests. If you are like me, I don’t really want to read all the first-person narrative that the AI delivers alongside the code. If we decrease how much narrative the AI generates, we reduce the number of tokens it uses.
Let me introduce what is being called “Caveman Mode”. In this case we instruct the AI to answer in as short a response as possible, except when urgent. We also drop the pleasantries and preambles to the chats.
There are lots of examples on GitHub, but here is one that I have been using for Business Central development.
name: skill-caveman-mode description: > Ultra-compressed communication mode. Cuts tokens ~75% while keeping technical accuracy. Use when user says “caveman mode”, “talk like caveman”, “use caveman”, “less tokens”, “be brief”, “/caveman”, or requests token efficiency. Supports intensity levels: lite, full (default), ultra. argument-hint: ‘lite | full | ultra’
Caveman Mode
Activated once, persists for session.
Switch Commands
| Command | Effect |
|---|---|
/caveman lite |
Short sentences, no filler, keep grammar |
/caveman full |
Fragments ok, drop articles, short words |
/caveman ultra |
Abbreviate everything, arrows (→), minimal words |
stop caveman / normal mode |
Off |
Default level: full
Rules (all levels)
- No filler (just/really/basically/etc)
- No pleasantries or hedging
- Keep tech exact
- No long sentences
- Prefer symbols:
→=
Response Pattern
[thing] [action] [reason]. [next step].
Level Examples
Prompt: Fix customer parse loop (inbound write triggers outbound).
- lite: “Trace trigger chain and locate loop point. Add integration context flag before inbound modify and consume it in OnAfterModify. Verify normal user edits still queue outbound updates.”
- full: “trace loop. inbound modify sets skip flag. OnAfterModify consumes flag once. user edit still queues.”
- ultra: “ind loop. set skip before inbound write. consume once on modify. user edit -> queue.”
Auto-Clarity Exceptions
Drop caveman for these; resume after:
- Security warnings
- Irreversible action confirmations
- Multi-step sequences where fragment order risks misread
Example:
Warning: This will permanently delete all rows in the
userstable and cannot be undone.Caveman resume. Verify backup first.
Scope
Code, commits, and PR descriptions: write normal regardless of level.
That’s it, a nice simple skill that reduces token usage by having a little less talk and a lot more action. Let me know if you have used Caveman Mode or intend to. Do you have any other tips on saving token costs? Share in the comments below.
Aardvark Labs 
Leave a comment