Source: raw/Claude_Code_-_Build_Your_First_AI_Agent_in_20_Minutes_100%_No_Coding.md (https://www.youtube.com/watch?v=F8jidmEkIeM), YouTube auto-captions fetched 2026-10-30. Creator: Grace Leung (youtube.com/@graceleungyl) Platform: YouTube Duration: 20:01 Published: 2026-10-30 Skill level: Beginner (“No coding, no terminals”; everything in the Claude desktop app)

Grace Leung, an AI-for-marketing educator who says she doesn’t code, builds one content-creator agent in the Code tab of the Claude desktop app. The agent turns her videos and newsletters into animated Instagram carousels. She takes it through five levels: workspace, agent, playbook skill, memory, then cloud runs on a schedule. The video is sponsored by Anthropic, and she says she is on the Max plan. It is recent (Opus 5.5 era), so the interface steps match the current desktop app.

Key Takeaways

  • An agent is judgment, not a chatbot. “A chatbot answers your question, but an agent takes the job off your plate.” Given a goal, it decides what to do first, which tools to use and whether the result is good enough.
  • Five parts of every agent: the brain (the model), the context (what it knows about your brand), the job (the goal), the rules (how it works and what it must never do) and the tools (what it may use).
  • Level 0, the workspace, is “the step most people skip”. The workspace is a folder (“Business OS”) holding her logo, personal photos and a few real Instagram posts as references. Claude read her website and those files, asked follow-up questions (which fonts, why she likes the example posts), then wrote CLAUDE.md plus a brand-context file covering business, audience, voice and visual style.
  • Prove the workspace with an A/B test. In a fresh session she asked for next week’s posts on “Claude Code for Marketers”. The plan used her channels, angles and colour palette. The same request in an empty folder was generic. “The only difference is the workspace.”
  • Level 1: let Claude write the agent. Her “agent designer” prompt (linked in the video description, not in the transcript) interviews you about the job and writes the agent file into an agents folder, covering the five parts. Call the agent with @. The first carousel draft was on brand but had missing images and basic, undersized type. “Very rarely is the first version of the agent … the final one.”
  • Level 2: a skill makes the agent consistent. Without one, the agent “makes its own calls on the hook, the layout” every run. Agent = teammate; skill = “the training manual”. She had Claude interview her (“think hard”) about energy and signature emotions, then saved a content playbook skill with a workflow, checklist criteria and motion. Run through the slash command, it animated every slide and exported each as its own video, ready to post.
  • Wire the skill in twice. Assign the skill to the agent, then update CLAUDE.md so Claude uses that agent “whenever I ask for social posts or carousels.” A new newsletter then came back as an animated carousel built with the playbook. You can follow it in the background-tasks view.
  • Level 3: memory makes it self-improving. “Every session is day one” by default. She asked Claude, in plain English, to create a memory file for the agent next to CLAUDE.md, to be read before every job. The key instruction: save only feedback that applies to future posts, not one-off edits. After feedback (no brown boxes, varied inner-slide backgrounds), lessons appeared in the file, and the next carousel avoided both faults.
  • The three files an agent reads: the onboarding guide (CLAUDE.md, “what you hand them on day one”), the playbook (the skill) and the memory (the notebook of lessons).
  • Level 4: put it in the cloud. Add a connector (Customize → Connectors → Notion) and tell the agent it can use it. Ask Claude to sync the folder into a private GitHub repo. From the Claude mobile app, start a Code session on that repo and say “check my Notion content inbox and turn those pending ideas into carousel posts.” It ran in the background with the phone screen off, returned Notion links and marked the inbox items done.
  • Then schedule it. Schedule → new task “Daily Content Drop”: every evening, check the Notion inbox, make carousels and propose playbook improvements from her feedback. Under advanced settings, pick the GitHub repo so it runs in the cloud. “Whatever I drop in my inbox during the day, the draft are waiting for my review the next morning.”

Try It

The five steps she gives at the end, as a checklist:

  1. Workspace. One folder; drop in your logo, a few real examples and your website URL. Ask Claude to write CLAUDE.md and a brand-context file, and answer its questions.
  2. Agent. Describe one job and let Claude interview you and write the agent file. Run it on a real input and iterate.
  3. Playbook skill. Have Claude interview you about what the first output got wrong, save the answers as a skill, assign it to the agent, and add a routing line to CLAUDE.md.
  4. Memory. Create a memory file next to CLAUDE.md. Tell the agent to read it before every job and to save only reusable lessons.
  5. Cloud. Push the folder to a private GitHub repo, connect your inbox tool, and create a scheduled cloud task.

“You don’t need to do it all at once. Just start with one level at a time.”

Implementation

Tool/Service: Claude desktop app (Code tab), Claude mobile app, Opus 5.5, the Notion connector, GitHub.

Setup (desktop choices she makes):

  • Local vs cloud. She starts local (“runs right on your computer”) and moves to cloud at level 4.
  • Permission mode: auto. It “let the agent work without interrupting you, but still blocks anything that is risky.” See Auto Mode.
  • Model: Opus 5.5, because it “follows detailed instruction precisely and has great design taste”. See Claude Opus 5.5.
Business OS/
  CLAUDE.md                 # onboarding guide; routes carousel requests to the agent
  content-memory.md         # lessons from feedback (her file sits next to CLAUDE.md)
  brand/brand-context.md    # business, audience, voice, visual style
  assets/  examples/        # logo, photos, real IG posts
  .claude/
    agents/content-creator.md
    skills/content-playbook/   # workflow, checklist, motion rules

The skill path (“under the Claude folder”) and the memory file’s location come from the video; the other file names are illustrative.

Memory instruction (paraphrased from the video):

Create a memory file for my content creator agent, next to CLAUDE.md.
Read it before every job. Only save feedback that applies to future
posts, not one-off edits.

Cost: Not given beyond “I’m on the Max plan myself.” The video is sponsored by Anthropic and links a sign-up page.

Integration notes:

  • What travels to the cloud. A cloud session clones the repo, so CLAUDE.md, skills and agents come along, but personal ~/.claude configuration does not (see Cloud Sessions Field Guide). Her memory file lives in the repo, which is why the cloud runs used it.
  • Memory written in the cloud. Cloud sessions work on their own branch. Lessons the agent saves during a cloud run reach your laptop copy only after that branch is merged; the video doesn’t show this step.
  • Built-in alternatives to a hand-made memory file. Claude Code has auto memory, and subagent definitions accept a memory: field (see Subagents). Her plain-file approach is easier to inspect and edit.
  • Scheduled cloud tasks are routines. Remote routines run with the laptop closed, have a one-hour minimum interval and clone the repo fresh on each run (see Claude Code Routines).

Open Questions

  • The “agent designer” prompt is only linked from the video description; its text isn’t in the transcript.
  • “GrooveAnchor” is the caption’s spelling of the website she onboards from; the real name isn’t confirmed.
  • How the animated slides were rendered (which library or tool the skill used) isn’t stated.
  • She doesn’t say whether the Notion connector needed extra setup to work in cloud sessions and routines. It worked in her demo.
  • Sponsored by Anthropic: the video shows no failures beyond styling faults, and it gives no run times or usage figures.