Our family talks to a local AI all day — from our phones, over Discord. It summarizes videos, answers from our notes, and it costs nothing a month. This week we taught it two brand-new abilities by writing two small text files, and both of those files are on this page for you to take.

What a “skill” actually is

Most people teach their AI by repeating themselves: every chat, the same instructions again. A skill is that instruction set written down once — a markdown file that says when to act, how to act, and what you’re never allowed to do. Our agent framework (Hermes — though the pattern works in any framework that reads instruction files, including Claude-style skill folders) loads a folder of these. Teaching it something new is not fine-tuning and not an API: write a file, drop it in the folder.

message (Discord) → local brain → skill file → local tools → answer

If you can write a recipe card, you can teach a local AI.

Download: yt-summarize.SKILL.md

The whole skill is ~30 lines. The mechanics: pull the video’s auto-subtitles with yt-dlp (no download, no API key), strip the timestamps, dedupe the rolling repeats, and summarize what the video says — not what its title implies. The rules matter more than the mechanics:

  • Never claim you watched the video. You read its transcript; say so if asked.
  • If no subtitles exist, say so honestly and offer the description instead.

The bug worth stealing our fix for

Our first summary called our presenter “Alex” — and invented a model name that doesn’t exist. YouTube’s auto-subtitles mangle proper nouns, and a summary inherits the mangling. The fix is five lines in the skill file: a known-name repair table (“Alex” → the correct name), and a hard rule — never invent a name to fill a garble; write “(name unclear in transcript)” instead. Re-ran it: fixed, permanently. That’s the whole loop: the AI makes a mistake, you edit a text file, it never makes that mistake again. Try that with a chatbot subscription.

Skill two: answer questions from my own notes

Download: my-notes-qa.SKILL.md

This is the one a cloud chatbot can’t offer without you uploading your life. The rules block is the entire value:

  • Search first — grep the notes folder; never read everything into context.
  • Answer only from what the notes say.
  • Cite the file the answer came from.
  • Never invent. If it’s not in the notes, say so.

We tested it with a trap: asked about a dentist appointment that doesn’t exist anywhere. It said so, listed what the notes do contain, and pointed out that the only “dentist” text on the whole machine was test data from an unrelated dataset. It refused to invent and it showed its work — that answer, not the fast ones, is what makes an assistant with file access trustworthy.

The GPU tax nobody warns you about

Our first live test of skill one took 64 minutes and 16 seconds. Nothing was broken — the graphics card was busy with image tools holding ~14 GB, and the model spilled into slow memory. Cleared the card, asked again: 2 minutes 19 seconds. Same question, 28× faster.

If your assistant shares a GPU with games or renders, this will happen to you. The rule is one sentence: your AI shares the stove with everything else you cook on — clear the stove before dinner. One command frees a resident Ollama model immediately:

curl http://localhost:11434/api/generate -d '{"model":"<your-model>","keep_alive":0}'

The honest bill

$0/month and the electricity. Two real taxes: an afternoon of setup, and patience — a local brain doing real work with tools takes a minute or two per answer, not two seconds. It’s a sharp colleague down the hall, not an instant genie. For “read this video” and “check my notes,” it’s already better than what we were paying for — because it’s ours, and because it accumulates: a chatbot forgets you when the tab closes; a skills folder compounds. Ours is at 109 files and counting.

Next in this series: giving the same assistant a voice in a Discord voice channel — a real competitor to realtime voice assistants, running entirely in your house.

Related: your first local AI with Ollama · make a local AI play Minecraft · what GPU do you need · can your PC run it