The integration works. The agent can fetch YouTube transcripts, query Notion, post to Discord. Step 6 of the §8.1 process — "save as a skill" — is the moment that turns a working chat into a permanent capability. Without it, the integration vanishes on context reset (usually overnight). With it, the integration is a permanent part of the agent's toolkit, auto-invoked when the right prompt comes in.

The channel's framing for why skills are the load-bearing pattern: the context window clears overnight, so anything stored only in chat memory is gone by morning. Skills survive because they live on disk, not in chat. The deeper claim from the second skills video: skills beat memory because only metadata loads at session start. The agent can ship hundreds of skills without bloating the prompt, and each skill is loaded on-demand only when needed. The "Matrix analogy" from the original skills video: Neo doesn't know kung fu all the time — only when he needs it. Skills are loaded the same way.

This article walks through what a saved skill looks like, where it lives, how to verify it's actually loaded, and the failure modes the channel flagged — including the load-bearing 40% context threshold ("dumb zone") that motivates saving skills in the first place.

What you'll learn

  • Skills are persistent recipe books that survive the nightly context wipe. They live at ~/.openclaw/skills/ (or ~/.claude/skills/ or a project-local .openclaw/skills/). When you save an API integration as a skill, the integration persists across sessions, reboots, and context resets.
  • Only metadata loads at session start. The agent reads the skill name, description, and trigger keywords at boot. The full skill content is loaded only when the agent decides the skill is relevant. This means you can have hundreds of skills with no context overhead — the host runs "hundreds or thousands of skills" in production (txTowBmYXMc, public.videos.summary_key_takeaways).
  • The 40% context threshold is the "dumb zone." Once context passes ~40% of the window, the agent starts ignoring instructions, returning wrong answers, and generally behaving erratically. The fix is to move every recurring workflow out of chat memory and into a skill. Each skill moved out of memory is context reclaimed for actual work.
  • You build skills by talking to the bot. The host's "presentation skill" was built by telling the agent his preferences and asking it to save them as a skill. The bot wrote the spec, structure, host location, and chained a deep research skill — none of it required hand-writing markdown. Refinement stays conversational: "update your skill on making presentations" or "I like light theme."
  • The skill-as-step-6 prompt is short. Once the integration works in chat, send: "save this as a skill named fetch_[service]_data so you can use it automatically in future conversations." The agent writes the skill file, and it's loaded as metadata from the next session.
  • The cron layer is the load-bearing piece for automations. A nightly rebuild, a morning briefing, a weekly report — these are skills triggered by cron, not by user prompts. The skill pattern from this article + the cron pattern from Course 1 §1.6 = the standard automation recipe.
  • The "clawhub virus" warning. The OpenClaw skill registry (clawhub) has had security issues. Use it for inspiration only — have the bot write the skill to your spec, not install someone else's. Someone else's skill encodes their preferences, not yours.

What a saved skill looks like

The skill is a markdown file with YAML frontmatter. The frontmatter is the metadata — name, description, tags, triggers. The body is the recipe the agent follows when the skill is invoked.

---
name: fetch_youtube_transcript
description: Fetches the transcript for a YouTube video using the YouTubeTranscript.io API
tags: [api, youtube, transcript, content]
trigger: youtube transcript, video transcript, get transcript
---

# Fetch YouTube Transcript