How to Use AI Agents for Total Beginners: A Crash Course w/ Agent Builder James McAulay

Main ideas & lessons

• AI agents are more than chatbots

  • A chatbot mostly responds.
  • An agent can take actions by:
    1. Using an LLM (“brain”)
    2. Having the right context (knowledge about you/business)
    3. Connecting to tools/software to actually complete work

• The #1 accelerator is context

  • Agents become useful/accurate when they can access your business + personal goals + records, rather than relying only on generic conversation or “automatic memory.”
  • The speaker contrasts:
    • “CFO with no data” → generic advice
    • “CFO with your accounts + access” → specific, actionable recommendations

• Build your “second brain” as a reusable context folder

  • Create a folder of static documents (often Markdown) that an agent can read every time.
  • This context typically includes:
    • Background research on the company/person
    • Your LinkedIn/profile content exported to files
    • Role profile/goals
    • A system instruction file (e.g., claude.md) that sets the agent’s behavior (e.g., “you are my chief of staff” and “read these files first”)

• Use MCP integrations so agents can plug into tools reliably

  • MCP (Model Context Protocol) is presented as a standard way for agents to interact with apps (e.g., Notion, Todoist).
  • MCP avoids each model/vendor building thousands of bespoke integrations.
  • With MCP, the workflow is:
    • Human/agent sends natural language
    • MCP layer translates to API calls
    • Tool returns results back in a usable form for the agent

• Use “skills” to package repeatable workflows

  • A skill is described as a saved, refined instruction bundle (like a large prompt) that you can run on demand.
  • Skills can be:
    • Triggered automatically when relevant
    • Iteratively refined so output quality is more consistent
    • Shared with teammates/customers
  • The speaker emphasizes quality comes from:
    • Giving lots of feedback during skill creation
    • Optionally simulating the skill across multiple past cases/personas to detect weaknesses

• Run tasks proactively (scheduled) without your computer

  • Scheduled tasks exist in two modes:
    • Local scheduled tasks: requires your computer on (access to your local files)
    • Cloud scheduled tasks: runs on vendor servers; your computer can be off, but it must rely on context available in connected apps (e.g., Notion), not local-only files
  • Cloud scheduled tasks are positioned as the “final form” to build a proactive chief of staff agent that runs in the background (e.g., daily summaries, end-of-day reports, weekly SEO scans).

• Agents scale along “proactivity levels”

  1. Information: summarizes what happened
  2. Suggestions: summarizes + recommends what to do next
  3. Draft work: summarizes + drafts emails/invoices/tasks for approval
  4. Self-improvement: reviews its own performance and updates instructions/skills (e.g., modifies claude.md, refines skills)

• Security is central

  • The key recommendation is permission hardening:
    • Check each connector’s permissions
    • Block destructive actions (e.g., delete tasks/events)
  • Privacy settings matter:
    • Personal accounts: turn off “help improve models”
    • Team/enterprise: typically off by default; also consider data residency constraints
  • Prompt injection risk:
    • Connecting agents directly to inboxes (e.g., Gmail) can let malicious emails influence the agent.
    • Mitigation: use a dedicated agent inbox (unknown to others) and forward/collect only what the agent should see.

• Portability: stop relying on “chat memory”

  • By using context folders, MCP connections, and skills, the workflow is more transferable between:
    • Different agent platforms/harnesses
    • Even different model providers (the “same context + skills” can work across systems)

---

Methodology / instructions (organized as practical steps)

A) Build a beginner-friendly “Chief of Staff” agent (context → instructions → tools → action)

Step 1: Define the agent’s role

• Create an instruction file (e.g., claude.md) that:
  • Assigns the agent a role (e.g., “chief of staff to Grant”)
  • Lists explicit goals for the agent to accomplish
  • Enforces workflow rules like:
    • “Read these files before answering”
    • “Be direct”
    • “Follow up on commitments”

Step 2: Create a “context folder”

• Interview yourself using a “skill” (the speaker mentions a skill called “Set Up My Second Brain”).
• Export/prepare:
  • Background research about the person/company
  • Profile/LinkedIn content as files (often Markdown)
  • Goals and working preferences
  • Any other documents the agent should consult

Step 3: Ensure context is actually loaded

• Use the platform behavior where the agent sees the folder on every run.
• Be aware that the agent may also incorporate built-in account “memory,” which can conflict with your files.

Step 4: Connect tools using MCP

• Add MCP-based connectors for the software you want the agent to act on (examples given):
  • Todoist (tasks)
  • Notion (documents)
  • Resend (sending emails)
  • CRM + calendar + meeting transcription tools (as examples in the sales workflow)

Step 5: Test a simple action

• Example provided: “Create a task in Todoist for follow-up tomorrow at 6pm”
• Verify:
  • Task creation works
  • Deadlines/labels/priority updates work
  • The agent is not performing unintended destructive actions

Step 6: Package repeatable tasks as skills

• Convert “end-of-day” or “transcript-to-tasks” behaviors into a skill:
  • Input: meeting transcripts (or other sources)
  • Output: commitments extracted + tasks created + email summaries drafted/sent
• Refine via feedback and iteration for consistency.

Step 7: Run proactively with scheduled tasks

• Choose scheduling mode:
  • Local: only if you can keep computer on and it needs access to local files
  • Cloud: if you want the computer off; store the context in connected cloud apps (e.g., Notion), not only local folders
• Create a cloud scheduled task:
  • Example concept: “End of day chief of staff report” every day at a fixed time
  • It pulls meeting transcripts from Notion
  • Uses skills to extract commitments and save tasks to Todoist
  • Sends a recap email (via Resend)

---

B) Security hardening checklist (permissions first)

• For each connector/plugin (e.g., Todoist, calendar, CRM):

  • Inspect what actions are enabled:
    • Read-only vs write vs delete
  • Block destructive tools:
    • Example: block Todoist “delete tasks”
  • Re-test with a high-risk prompt:
    • Attempt “delete all tasks”
    • Confirm the tool is blocked and the agent refuses/fails safely

• Inbox / email safety:

  • Do not let an agent read arbitrary inbound email directly.
  • Use a dedicated agent inbox:
    • Unknown to outsiders
    • Only forwarded/BCC’d relevant messages

• Model training/privacy controls:

  • Personal accounts: disable “help improve models”
  • Team/enterprise: off by default (plus consider region/data residency constraints if required)

---

C) Skill creation quality-control method (“quality floor”)

• During skill creation:

  • Provide as much feedback as possible
  • Iteratively adjust:
    • Too short / too long
    • Which details to include/exclude
    • Formatting expectations (bullets vs paragraphs, etc.)

• After skill is drafted:

  • Optionally run simulations with sub-agents:
    • “Simulate 10 different people using this skill over past weeks”
    • Have the master agent critique results
    • Identify where the skill fails or becomes vague

• Operationally:

  • Treat skill building like engineering:
    • Start using it
    • Update skills when you spot errors (“never make the same mistake again”)

---

Speakers / sources mentioned

People (speakers/guests)

• Grant (host; unnamed in subtitles)
• James McAulay (guest; also described as founder of Agent Accelerator; the video title includes “James McAulay”)
• Corey (mentioned as absent due to vacation)
• Dr. Elena Zoo (Intel interview mentioned)
• Schaefer Twins (asked a security question)
• Evan (asked multiple questions, including scheduling/triggers and context)
• Learn Promptly AI (asked a question about customizing instructions for account skills)
• Debbie (asked a question about capabilities/agent access—context unclear)
• Mobility FL (asked questions about transferability, agent model selection, token management)
• G Hogarth (made a correction about Markdown/Google Drive capabilities)
• Aber Aarian (asked about using Claude for ads with Google Ads/Meta)

Organizations / products / platforms (sources/tools)

• Neuron AI live (event/brand mentioned)
• Claude / Claude Code / Claude Cowork / Cloud scheduled tasks (Anthropic ecosystem; also “Claude MD” referenced)
• ChatGPT / OpenAI ecosystem
• OpenRouter
• Anthropic
• MCP (Model Context Protocol)
• Notion
• Todoist
• Resend (email sending)
• Google Drive / Google Docs / Gmail
• Zoom / Google Meet (meeting transcription mentioned)
• WhisperFlow / Whisper-style transcription
• GitHub / Cloud Code
• Fable, Opus, Sonnet, Haiku (model families mentioned)
• Glean (enterprise search tool mentioned)
• skills.sh (skills repository mentioned)
• AgentMail / Agent mail inbox (mentioned for secure inbox routing)
• Agent Accelerator / agentacelerator.ai (speaker’s program/course)
• Eve framework and Vercel (mentioned as part of an agent harness discussion)