Video summary

AI Agents: The Most Valuable Skill You Can Learn in 2026 (Full Course)

Main summary

Key takeaways

Educational

Main ideas & lessons

  • AI is shifting from “chat” to “agents.”

    • Stage 1 (chat): question → answer (e.g., ChatGPT / Claude / Gemini) Best for small, quick questions and guidance.

    • Stage 2 (agents): goal → result Agents plan, execute steps, use tools, and deliver finished outputs.

    • Using agents can make people 5–10x (or ~7x mid estimate) more productive, widening the gap between teams that adopt agents vs. those that stick to chat.

  • Agents are like hiring a capable employee—onboarding matters.

    • To make an agent actually useful, you must supply:
      • Context (who you are / your business / your voice)
      • Tools (Gmail, Slack, browsers, scrapers, etc.)
      • Skills (your SOPs/processes—how you want tasks done)

  • The “agent loop” under the hood:

    • Agents repeatedly perform: Observe → Think → Act
    • They continue looping until the task is complete (as defined by the goal / “what done looks like”), then output the result.

  • “Agents harnesses” and “MCPs” are key enablers.

    • The platforms that run the observe-think-act loop (e.g., Claude Code, Codeex, etc.) are described as agent harnesses.
    • MCP (Model Context Protocol) is presented as the standard way to connect agents to external tools (Gmail, Slack, Calendar, etc.) without heavy custom integration.

  • Build context with markdown files you own (not black-box chat memory).

    • Use markdown assets so the agent can read and you can control what it knows.
    • Prefer lean/clear context to avoid consuming too much of the context window.

  • Folder-based “Personal AI OS” / operating system concept

    • Instead of many separate “agents,” the recommendation is to run from one OS-style folder and use:
      • a global northstar file (e.g., claude.md / agents.md)
      • subfolders for different pillars/projects
      • skills stored in a dedicated hidden folder (e.g., .claude/skills)
    • This scales as your needs grow and supports global vs project-level context.

  • Skills = SOPs for AI (repeatable, non-sloppy execution).

    • Skills are packaged procedures with:
      • name
      • description
      • contents (step-by-step instructions, tool usage, formatting rules, examples)
    • Skills enable repeatability across sessions and reduce re-explaining preferences.
    • Skills can be iterated (V1 → V2 → … “sanding down” to match your taste).
    • Skills can be chained/orchestrated (one “workflow/orchestrator” skill calls multiple other skills in sequence).

  • Must-have tooling for the “Act” step (suggested).

    • The speaker names several categories of tools/MCPs as especially useful:
      • Appify/Ampify (marketplace of scrapers for many data sources)
      • Firecrawl (reads/extracts from websites, beyond simple web search)
      • Composeio (connector hub; simplifies managing multiple tool connections)
      • Browser-control MCPs: Chrome DevTools, Playwright
      • Higsfield (image/video generation via an MCP wrapper)
      • Mentions also: Meta ads-related integration, 1Password access (caution)

  • Bottom-up AI adoption for companies

    • Best path to becoming “AI native” is not only top-down specialists building for employees.
    • Instead: train employees to build their own AI OS / skills, making them “100x employees,” then automate more work from the bottom up.
    • Companies should make AI adoption part of culture (sharing weeks, competitions, skill-of-the-month).

Methodology / instructions (detailed checklist)

A) Build an agent that’s actually useful (onboarding framework)

Use Context + Tools + Skills.

Context (before tasks)

Create markdown files/assets that describe:

  • About you / your story / background
  • Business info
  • Brand voice
  • Offer catalog (services, pricing, upsells/downells)
  • Ideal customer profile
  • Rules/preferences (what “good” looks like)

Recommendations:

  • Store as markdown (easier for agents to ingest than PDFs)
  • Keep it lean to preserve context window capacity
  • Prefer objective, non-subjective statements to reduce “AI slop”

Suggested “northstar” file:

  • A file like claude.md / agents.md that autoloads every new session.

Tools (Act step capabilities)

Connect external systems the agent needs to perform work:

  • communication: Gmail, Slack
  • planning: Calendar
  • data/task management: Notion, ClickUp
  • web extraction: scrapers, web page readers
  • automation / booking: Cal.com

Use MCP / connectors to enable tool access consistently.

Skills (how you want work done)

Convert repeatable workflows into skills (SOP-like markdown).

A skill should typically include:

  • Name + description
  • Step-by-step procedure (what to do first, second, third)
  • Tool instructions (what to scrape/search/use)
  • Output formatting rules
  • Quality checks / acceptance criteria

Benefits:

  • Reduces back-and-forth
  • Preserves your “taste” across sessions
  • Makes workflows repeatable and automatable

B) Agent loop definition (what happens while it works)

When you give a goal:

  • Repeatedly run:
    • Observe: look at available data/context/files/tools
    • Think: decide next action based on the goal and what’s missing
    • Act: use tools/commands to proceed

Stop criteria:

  • Determined by the goal’s specificity (“done looks like X”).
  • If vague, it may stop at a more subjective interpretation.

C) Build your “Personal AI OS” folder structure (macro → project)

High-level organization approach:

  • Create a top-level folder like OS/ (holding company concept)
  • Inside it:
    • global pillars/projects (e.g., Open Residency, Iconic, Personal)
    • each pillar contains subfolders (content/newsletter/website/marketing/etc.)

Add a global northstar context file (autoloaded) such as:

  • claude.md at the OS level

Add project-specific northstar overrides:

  • project folders can include their own claude.md to specialize behavior

Use tagging conventions:

  • use slash/at style commands (e.g., @marketing) so the agent knows which folder scope to use.

D) Create skills (two approaches)

  • Goal-first
    • Ask the agent to create a skill directly given a desired outcome and inputs (e.g., “Build a brand guidelines skill using this PDF”).
  • Process-first (preferred)
    • Run the workflow once end-to-end.
    • After you’re satisfied, instruct the agent to “package this process into a skill”.
    • This yields your “V1 skill,” which you refine iteratively.

E) Skill anatomy (what a skill file contains)

  • Each skill has:
    • skill.md
      • name
      • description
      • contents

Behavior:

  • In sessions, the agent initially loads name/description (low context cost).
  • When the task requires a specific skill, it loads the full contents (progressive disclosure).

F) Practical automation suggestion example (agentized workday)

  • Setup goal:
    • connect tools + context
    • write a task like: “Find X, analyze it, produce Y report, deliver in PDF/HTML, and schedule it daily”
  • Then:
    • on a schedule, run the agent to perform repetitive work (reports, outreach drafting, inbox triage, etc.)

Speakers / sources featured (identified)

  • Remy (main speaker; founder of “AI with Remy” / “Open Residency” context; also referred to as “Remy” and “Rem Dog”)
  • Ollie Leman (mentioned as origin of a Claude council skill; also “Ollie” as a friend whose ads were used in the demo—identified by name)
  • Chris Voss (negotiation expert; referenced for a “language/prompting” inspiration)
  • Ross Mike (referenced as an explainer for the MCP “translator” concept)
  • Uncle Greg (mentioned in context of Higsfield / demos)
  • “Ask Cat GPT” (creator referenced as a content inspiration)
  • Momentous (sponsor)
  • Ketone IQ (sponsor)
  • OpenAI / Anthropic (mentioned as LLM providers; not featured as individuals)
  • Google (mentioned via Chrome DevTools and general context; not featured as an individual)
  • Meta (mentioned via ads library/integration)
  • Composeio / Claude / Cal.com / ClickUp / Notion / Gmail / Slack / Stripe / WhisperFlow / Higsfield / Firecrawl / Ampify/Appify (tools/platforms referenced as sources of capability, not speakers)

Original video