Caveman
Make your agent talk like a caveman. Same answers, far fewer output tokens, full technical accuracy. Brain big, mouth small - it shrinks what the agent says, never what it knows. Code, commands, and errors stay byte-exact. Its companion skills load dynamically, straight from the caveman GitHub repo at runtime - never baked into the prompt.
Why use many token when few do trick. This blueprint is one idea, done well: an agent that answers like a smart caveman - it drops the filler, the articles, the pleasantries, and the hedging, and keeps every bit of technical substance. Same fix, a third of the words, nothing lost.
The persona is a backstory - the way of speaking is the whole point, and it travels because it is a style, not a piece of infrastructure. But its extra skills - caveman-commit, caveman-review, caveman-compress - are deliberately NOT stuffed into that backstory. They live in a public GitHub repo and load dynamically, straight from that repo at runtime: a Caveman Skills skillset gives it two tools - list the skills in the repo, then fetch a skill's SKILL.md by path - so it pulls a skill directly from GitHub only when a task calls for one, keeping the prompt small. Point those two abilities at your own repo to swap in your own skills.
The backstory carries three things worth calling out. First, a hard preservation boundary: code blocks, shell commands, API and function names, commit-type keywords, and exact error strings are copied verbatim
- compression only ever touches natural-language prose, never the parts a reader will paste or run. Second, intensity levels - lite keeps full sentences but drops the fluff, full is classic caveman, and ultra strips every non-load-bearing word - so the same agent can dial terseness up or down per conversation. Third, an auto-clarity rule: it steps back into full, careful sentences for security warnings, destructive-action confirmations, and any multi-step sequence where a dropped conjunction could be misread, then resumes caveman once the risky part is past.
A note on the honest version of the pitch: this shrinks OUTPUT tokens, not input or reasoning tokens, so the real win is readability and speed as much as cost. It also keeps the user's language - write to it in Portuguese and it grunts back in Portuguese - because it compresses style, never meaning.
Extend it by wiring the same backstory into a channel integration (Slack, email, WhatsApp) so the terse voice shows up wherever your team already talks, or pair it with a dataset so grounded answers come back just as short.
Backstory
Common information about the bot's experience, skills and personality. For more information, see the Backstory documentation.
Skillset
This example uses a dedicated Skillset. Skillsets are collections of abilities that can be used to create a bot with a specific set of functions and features it can perform.
List Caveman Skills
List the skills available in the caveman repository - returns the name, description, and path for each.Fetch Caveman Skill File
Fetch a file from the caveman repository by path - use a SKILL.md path from the list ability to load a skill.
Terraform Code
This blueprint can be deployed using Terraform, enabling infrastructure-as-code management of your ChatBotKit resources. Use the code below to recreate this example in your own environment.
A dedicated team of experts is available to help you create your perfect chatbot. Reach out via or chat for more information.
