SKILLs Aren't Token Waste: How Custom SKILL.md Files Teach Agents Your Workflow
ievkz · reddit · 2026-10-26
The author went from dismissing SKILLs as token-hungry overhead to a convert after writing his own, and lays out four advantages over MCP:
- Tiny context cost: each SKILL adds only 100–200 tokens (name + description), so you can mount 30 at once; the agent loads full instructions on demand, or you invoke it explicitly.
- Covering obscure tools: models know Git and Python, but your in-house or closed-source program has no docs — a SKILL.md describing it saves the agent from hunting the web.
- Encoding business processes: agents know how to run tests and build, but not how your company ships; a SKILL can encode the release sequence (check diff → tests → CHANGELOG → bump version → build → tag → publish).
- Instructions + tools in one package: a SKILL folder can ship scripts, references, and templates that stay out of context until needed — agents can run ready-made scripts instead of burning tokens generating code.
The author's realization came while building blueprint4llm, prompted by LLM-generated plan.md walls of text he wanted replaced with shorter outputs containing flowcharts, tables, and UI prototypes.
Related event: Why Developers Are Embracing Custom SKILL.md Files for Agents(2 posts)→
More from coding & agent
- Compound rewrites Excel, PowerPoint and Word engines in Rust for AI agents — peterjliu · 2026-10-26
- Dev vibecodes open-source suite: guided LoRA trainer, dataset captioner, image tournament tool — Speedyrulz · 2026-10-26
- Thorsten Ball publishes long-form guide on how he works with coding agents — vivekhaldar · 2026-10-26
- MCP explained: the open protocol standardizing how AI agents connect to tools — robertonappe · 2026-10-26
- Seeed Studio open-sources reBot-DevArm: a 6-DoF robotic arm with sub-0.2mm repeatability — lukas_m_ziegler · 2026-10-26
- Dev Predicts Personal Agents Will Replace Operating Systems, Generating UI on Demand — evielync · 2026-10-26