--- title: "Does the Caveman skill actually save money? Sometimes" description: "The Caveman skill is a rule file. It tells a coding agent to answer in fewer" canonical: https://caveman.so/news/does-the-caveman-skill-save-money last-updated: 2026-09-15 --- # Does the Caveman skill actually save money? Sometimes The Caveman skill is a rule file. It tells a coding agent to answer in fewer words and to leave code, commands, paths and error messages exactly as they are. It has more than 100,000 GitHub stars and a lot of people install it expecting a smaller bill. Sometimes they get one. Sometimes they do not. This page says which is which, with the numbers, so you can decide in a few minutes whether to keep it on. ## What it does and does not touch The skill changes output. It does not compress input, context, files, or thinking tokens. Those are the job of the separate local proxy. If your bill is mostly input, which in agent work it usually is, the skill addresses the smaller counter. ## The measured win Ten ordinary coding prompts through the real Claude API, same model, with the skill and without. Output tokens per reply: | Task | Normal | Caveman | Saved | | --- | ---: | ---: | ---: | | Implement React error boundary | 3,454 | 456 | 87 percent | | Set up PostgreSQL connection pool | 2,347 | 380 | 84 percent | | Explain git rebase vs merge | 702 | 292 | 58 percent | | Refactor callback to async/await | 387 | 301 | 22 percent | | Average across all ten | 1,214 | 294 | 65 percent | The best and worst rows are shown on purpose. The skill wins big when the agent would otherwise write an essay. It barely wins when the answer was already mostly code. The table regenerates from the repository's benchmark script with an Anthropic key. ## The cost you pay every turn The skill's rules are about 5 KB and are injected into context on every turn, along with skill-list entries. That is roughly 1,000 to 1,500 input tokens per turn, paid whether or not the reply was going to be long. So the arithmetic per turn is: output tokens saved times the output price, minus about 1,250 input tokens times the input price. On a verbose agent that is comfortably positive. On a terse one it can be negative. ## Where it loses **Already terse work.** One user measured a net loss on short coding questions, where the fixed rule cost exceeded any output saved. If your sessions are quick lookups, expect this. **Per-request pricing.** GitHub Copilot charges premium requests. A shorter answer is the same request. The skill cannot lower a Copilot bill, and the same applies to any plan that meters messages rather than tokens. **Tool-side counters that disagree.** One Cursor A/B reported 4.3 million tokens with the skill against 1 million without, and twice the wall-clock time. We could not reproduce the run. The safe reading is that rule re-injection, retries, and cache or context accounting can overwhelm output savings in some harnesses, and that a session total can move in the opposite direction from the output-only table. **Whole sessions.** Prompts, context, files and injected rules all consume tokens. Session totals land below the 65 percent figure, and the only number that settles it is the provider's own bill on the same task with and without. ## When it wins - Long, explanatory replies. Architecture questions, debugging walkthroughs, anything where the model narrates. - Long sessions with verbose agents, where per-reply savings accumulate against a fixed per-turn cost. - When you value the shorter reply itself. Faster to read, less to scroll. The discount is the bonus, and for many people the readability is the product. ## Check it yourself in ten minutes 1. Pick one task you do often. Run it with the skill and without, three times each if you can. 2. Compare provider-billed totals, not a local token counter. In Claude Code that is `/cost`; elsewhere it is the provider's usage page. 3. If the fixed overhead beats the output reduction, turn it off for that workload. `/caveman off` in the session, or uninstall. ```bash caveman trial -- claude caveman trial report ``` `caveman trial` runs the A/B for you against a real session and prints the comparison. If your result contradicts the table above, open an issue and we will add it to the honest numbers page. ## Questions people ask ### How much does the Caveman skill save? On ten measured coding prompts, 65 percent fewer output tokens on average, from 1,214 to 294 per reply. That is output only. Whole-session savings are lower because the skill adds roughly 1,000 to 1,500 input tokens per turn and does not touch input. ### Can the Caveman skill cost more than it saves? Yes. On short, already terse work the per-turn rule cost can exceed the output saved. On per-request pricing such as Copilot premium requests it saves nothing. One unreproduced Cursor A/B went the wrong way by a wide margin. Turn it off where your own measurement is negative. ### Does it change the code the agent writes? No. Code, commands, file paths and exact error messages are never shortened. Only the prose around them is. ### Does it work with agents other than Claude Code? Yes. It installs into more than thirty agents including Codex CLI, Gemini CLI, Cursor, Windsurf and Cline. Some load it automatically, some need `/caveman` per session. The install table in the repository lists each. ### Is it free? Yes. The skill is MIT licensed and needs no account. The local proxy is a separate component with a different licence. ## Sources - [Honest numbers](https://github.com/JuliusBrussee/caveman/blob/main/docs/HONEST-NUMBERS.md), the losing cases with issue links. - [Benchmark table and script](https://github.com/JuliusBrussee/caveman#the-numbers). - [Install table](https://github.com/JuliusBrussee/caveman/blob/main/INSTALL.md).