← sources

I Thought SKILLs Were a Waste of Tokens Until I Wrote My Own

source: reddit (reddit:ollama)published 2026-10-25 09:44ingested 2026-10-26 02:39

retained text

I'll be honest, I didn't understand why LLM agents need SKILLs. MCP was clear to me: it's a useful thing that lets you give an LLM access to a database or to Figma. The only downside is that, to avoid spending too many tokens, it's better not to connect 30 MCP servers at once. And in general, it's better to turn off the unnecessary ones and connect only those that are really needed for the task. In this light, SKILLs seemed like useless overhead to me, eating tokens like crazy. On TikTok, I often saw tales along the lines of: connect this SKILL, or a whole bunch of them, and happiness is yours. My first thought was, if I connect them all, my subscription limits will run out not by lunchtime, but already in the morning.

https://preview.redd.it/qt3gjp583lxh1.png?width=837&format=png&auto=webp&s=ee8372347b1420b69d13ccc9cf8457293c6047d6

So why are SKILLs useful and needed after all? Well, first, compared to MCP, SKILLs do not inflate the context: each connected SKILL adds not tens of thousands of tokens, but one or two hundred. Since only the name and description of a SKILL go into the context at first, you can add 30 of your own SKILLs without much pain. When it becomes necessary to use a SKILL, the LLM agent will load its description on its own, if it understands from the task context that this is needed, or the user can simply tell it to use this SKILL for the job.

https://preview.redd.it/i597mp993lxh1.png?width=590&format=png&auto=webp&s=19e5f0b7f969320113c19ddb6ab4f29b5456e8ef

Secondly, LLM models are trained on the most popular and well-known things. For example, everybody knows that an LLM can work with Git without any problems or write a program in Python. But if you developed a program yourself, or your work requires some little-known tool or technology, then a SKILL is simply irreplaceable. It is enough to create a SKILL.md file and describe how to work with this program or technology. This makes life easier for the LLM: while doing your tasks, it will not have to search the internet for documentation and wander around websites to understand how to use it. And especially if there is nothing on the internet, because your program is closed to public access.

https://preview.redd.it/cx8gogrm3lxh1.png?width=690&format=png&auto=webp&s=8ef3977f3829887eef7f6b4563a420e385b44361

Thirdly, an LLM model usually does not know how your development process is organized. For example, an agent knows how to use Git, run tests, build a project, and create releases. You do not need to teach it these things. But it does not know how releases are usually made in your company. A Skill can describe the sequence of actions: check the changes, run tests, update CHANGELOG.md, increase the version, build a package, create a Git tag, and publish a release. So, a Skill helps the agent learn not a technology, but a specific business process.

https://preview.redd.it/a88wwkmo3lxh1.png?width=549&format=png&auto=webp&s=757d7e23ffe49a2e2c72bdef9a75832793949674

Fourthly, a SKILL is not just an instruction, but a whole set of tools. Besides SKILL.md, you can put ready-made scripts, documentation, templates, and other files into it. And the most interesting thing is, these files are not loaded into the context automatically. The agent reads them only when there is a need, and can simply run the ready-made scripts without spending tokens on generating code. So, a SKILL lets you combine instructions, documentation, and executable code in one package without inflating the LLM context. For example, the structure can look like this:

    my-skill/
    ├── SKILL.md
    ├── scripts/
    │   ├── generate.py
    │   └── validate.py
    ├── references/
    │   └── api.md
    └── assets/
        └── template.html

https://preview.redd.it/4hma9xjq3lxh1.png?width=695&format=png&auto=webp&s=7f48aec25914d8f681815d40babc3b0da98822a7

The funniest thing is, I started to appreciate SKILLs exactly when I wrote my own program, blueprint4llm. The idea for creating it came after I asked an LLM agent to generate a plan.md file describing a task. Usually it produces a huge wall of text, which you have to read slowly and with sadness. I wanted to make the text shorter and add more flowcharts, tables, diagrams, and user interface prototypes. As we know, in .md (Markdown) files the only visualization tools available are tables and Mermaid diagrams. There are no charts, plots, histograms, full UML diagrams, or UI prototypes. In the end, with blueprint4llm, I added support for formats such as PlantUML, draw.io, Pencil, Balsamiq, Graphviz, Figma, OpenAPI, Excalidraw, Vega-Lite, ECharts, and D2 to .md files.

https://preview.redd.it/wh0h3jvs3lxh1.png?width=708&format=png&auto=webp&s=3d080ab388906d6ca0d5ba1df666d48d1424caf8

With the blueprint4llm SKILL connected, you will get plan.md not as a kilometer-long text, but as flowcharts, prototypes, charts, and so on. By the way, blueprint4llm does not have to be used only for creating plan.md. For example, I used this SKILL to add an illustration to every paragraph of the article and make it more fun to read. This SKILL is also useful when you ask an agent to put together a full dossier on some company, technology, or person and format the resulting result.md file with blueprint4llm. You can then convert it to HTML or PDF, or ask the agent to show it right away in a browser with charts.

https://preview.redd.it/d4dvf7cu3lxh1.png?width=944&format=png&auto=webp&s=65e65c63f052d87d87283fcd10fc89563d69ce89

This is how you can install the blueprint4llm SKILL.

    # Codex
    npx skills add evgenyigumnov/blueprint4llm --skill blueprint4llm --agent codex
    
    # Claude Code
    npx skills add evgenyigumnov/blueprint4llm --skill blueprint4llm --agent claude-code
    
    # Pi
    npx skills add evgenyigumnov/blueprint4llm --skill blueprint4llm --agent pi

https://preview.redd.it/3zw2bmmx3lxh1.png?width=887&format=png&auto=webp&s=e4def8733724b80116c7e13dba4c03ed20637121

If you are interested in looking at the source code of blueprint4llm, the tool that generates all this beauty, or sending a pull request, welcome: [https://github.com/evgenyigumnov/blueprint4llm](https://github.com/evgenyigumnov/blueprint4llm)

https://preview.redd.it/n7ym2gimwlxh1.png?width=690&format=png&auto=webp&s=dfb99ace7e0580935026df6cccd5535053ebf5bb

document fetched 2026-10-26 00:37 · canonical https://reddit.com/r/ollama/comments/1xey67z/i_thought_skills_were_a_waste_of_tokens_until_i