Token-Saving Skills Actually Only Save 8.5%
dotey · x · 2026-07-13
This deep dive dubs token-saving Skills as "Telegraph-style Skills": much like paying per word in the old days of telegrams, people compressed messages to the bare minimum. Today, skills in certain AI coding tools attempt a similar approach, using shorter, rigid prompts to cut output tokens.
The post highlights the GitHub project Caveman: the creator injects a "caveman" command into tools like Claude Code and Codex, stripping out articles, pleasantries, and conjunctions to leave only technical essentials. Its README claims a 65% token saving, propelling it to the #1 spot on Trending in three days with 87,000 stars.
However, JetBrains offers a more conservative take from their testing. They ran controlled experiments on Claude Code across SkillsBench's 86 real-world coding tasks—using identical tasks, models, and budgets—for roughly 240 billed trials costing $106. The results showed that even under ideal conditions forcing Caveman to activate every time, it only saved 8.5% of output tokens.
The post concludes that:
- The advertised 65% token savings mostly apply to chat scenarios
- Savings in actual coding tasks fall far short of the hype
- Such "minimalist prompt skills" are likely a passing phase with limited long-term value
More from coding & agent
- Travel MCP Server adds flight, hotel, weather and budget tools for agents — modelcontextprotocol · 2026-07-21
- OpenAI hackathon project stalls as Codex struggles on voice, while Claude spots the issue — ColleenMBrady · 2026-07-21
- Douyin Video Analysis MCP turns share links into structured video summaries — modelcontextprotocol · 2026-07-21
- An indie builder open-sourced 50+ AI apps and says users should only pay for tokens — matchaman11 · 2026-07-21
- Shared agent workspaces fail in a fixed order, from stale reads to zombie writes — mrvladp · 2026-07-21
- pi 0.81.0 adds first-class integration with llama.cpp server — huggingface · 2026-07-21