Caveman and Ponytail: Two Free Skills That Cut Your AI Token Bill
Why your AI bill is mostly small talk
If you run agents all day, you already know the pain. We use Hermes and Claude Code on our team, and the token usage adds up fast. A big chunk of that spend is not the actual work. It is the agent narrating the work. "Hey guys, I just finished building the server, now I'm starting this up..." That kind of filler burns tokens and burns your time reading it.
Two skills we recently started using attack exactly this problem. They are called Caveman and Ponytail. They started as developer jokes on GitHub, but they genuinely work, and we now keep them in our setup.
Caveman: make your agent talk like a caveman
Caveman forces your AI to speak in short, blunt phrases. Think "rock smash good" instead of "that rock looks really big, let's find a way to smash it." The transcript example is telling: the same status update drops from 69 tokens to 19 tokens.
The advertised token reduction is around 65%, but treat that as a ceiling. Real savings depend on the task. Even a 30% cut matters when you are close to your quota at the end of the month. There is a second benefit people overlook: less reading. When the agent skips the friendly walkthrough and just gives you the key information, you move faster. If you are a beginner, the verbose mode is actually helpful. Once you know what you are doing, it is dead weight.
Installation is easy. Just ask your CLI agent, including Hermes, to install the caveman skill. Then restart your client and type /caveman.
Two quirks to know. First, it does not always auto-start, so you need to enable it for every session. Second, after a context compaction, the agent forgets it is a caveman. You have to re-run /caveman after each compact or it drifts back to polite essay mode.
Ponytail: make your agent lazy on purpose
Ponytail comes from the same philosophy, but it targets the work itself instead of the words. It tells the AI to do the laziest, simplest thing possible. No overthinking, no exploring five approaches, no conversation about the task. Just do it.
This shines on routine tasks. When we say "start the server" on a fresh Windows install, we do not want a guided tour of everything happening. We want the server running, the fastest and easiest way. There is a real difference between easy and simple here. Typing "start the server" is easy for you, but left alone the AI often picks a complicated path and keeps checking in. Ponytail cuts that off.
The key point: it still does the work. Lazy does not mean sloppy. Sometimes the laziest path is the simplest one, and the simplest one is the best one. Click the button, start the server, done.
The practical call
Install both. Caveman trims the output tokens, Ponytail trims the reasoning and back-and-forth. Together they cut cost and speed up your loop. Just build the habit of re-enabling Caveman after every session and every compaction, or you will wonder why your agent got chatty again.
We collect the skills our team actually uses on our Super Bash Learn website, along with a model tier list and visual benchmarks if you are still deciding which model to run. Check the skills section there if you want more like these two.