Cached at:
04/22/26, 09:07 AM
TL;DR: 100K-subscriber YouTuber Grace Leung breaks down her 2026 AI marketing stack and argues that “system thinking” beats chasing the newest tools.
## The Core Stack: 80% Claude, 20% Everything Else
Grace Leung’s daily workflow runs almost entirely inside Claude.
- She uses both Claude Web and Claude Code, pushing more tasks into the latter because it accepts connectors, custom skills, and repeatable markdown-based playbooks.
- For quick research she opens Perplexity or any frontier LLM, but the heavy lifting—briefs, scripts, audience analysis—stays in Claude.
Ryan Doser’s observation: high-performing marketers are migrating from “ChatGPT + 17 browser tabs” to local IDE/terminal setups where every prompt, skill file, and output lives in version control. Grace agrees, but adds a warning:
> “People treat Claude Code like a fancier chat bot. The tool isn’t the ceiling—your system is.”
## System Thinking > Tool Hopping
Grace’s rule of thumb: squeeze every ounce of value from the tool you already pay for before switching.
- Document reusable components: system prompts, skill markdown files, SOP templates.
- Build so the workflow survives if you step away.
- Only migrate when you hit a hard ceiling, not because a TikTok told you “Claude is dead, Gemini is king.”
## Consolidation Wave: One Model, Many Use-Cases
Two years ago marketers juggled a dozen point solutions—copy generators, thumbnail AIs, landing-page builders.
Today Claude, Gemini, or GPT-4o can handle:
- Long-form articles
- Slide decks
- Simple motion graphics
- HTML/JS landing pages
The result: “SaaS graveyard” headlines appear weekly as frontier models subspecialize.
## Enterprise vs. Solo Creator: Landmines to Avoid
### Team Collaboration
- Corporate data sits in Slack, Drive, SharePoint. Any AI layer must plug into those spaces.
- Some teams now treat GitHub as the collaboration hub so Claude Code can natively read issues and PRs.
### Change Management
- Pick 1–2 “AI champions,” give them sandbox budget, and make them write the internal playbook.
- Resist executive pressure to “automate 1,000 workflows by Friday”; existing staff already have KPIs and zero risk tolerance.
### Compliance & Cost
- API calls are cheap until they’re not; one misconfigured agent can burn five-figure credits overnight.
- EU data rules, HIPAA, or SOC-2 can instantly veto Claude. If Gemini is already approved inside Google Workspace, build there first.
## Grace’s YouTube Production Loop
1. **Ideation Gap**
- Feed Claude her past 50 videos + competitor metadata.
- Ask for “angles my channel hasn’t covered that my ICP still Googles.”
2. **Outline & Hook Writing**
- Claude returns three hooks, timestamps, and transition sentences.
- Grace rewrites 30–40% to inject personal stories—what she calls “information gain that algorithms can’t fake.”
3. **ICP Reality Check**
- A second Claude skill reviews the script from the Ideal Customer Profile viewpoint: “Where will they drop off? Which line triggers trust?”
4. **Notebook LM Assembly**
- Sources (PDFs, URLs, raw notes) go into Notebook LM → clean outline or data table.
- Export to Google Docs, loop back into Claude for slide copy or newsletter version.
5. **Publishing & Repurpose**
- Final metadata (title, description, chapters) generated by Claude, approved manually.
- Audio cleaned in Descript (AI silence removal), thumbnail still made in Figma—”because eye-tracking nuance matters.”
She rarely lets AI draft an entire video; voice-over cadence and on-camera trust remain human levers for conversion.
## AI-Powered SEO: Audience Insight, Not Keyword Spam
Grace’s stance:
- Keyword research is still useful for demand sensing, but Google’s helpful-content update rewards “information gain” (unique data, first-hand experience).
- Instead of SERP scraping → auto-articles, she:
- Uses Appify to scrape Reddit threads, YouTube comments, and X posts.
- Lets Claude cluster pain points into content briefs.
- Films or writes the solution, adding proprietary data or personal case studies.
Result: pages index faster, earn backlinks, and survive algorithm updates.
## Quick Wins You Can Implement Today
1. Pick one repetitive task (weekly newsletter, landing-page QA, lead-list cleaning).
2. Write a one-page SOP in plain English.
3. Convert the SOP into a Claude skill markdown file (context, input format, desired output).
4. Run 3–5 tests, refine, then share the skill with teammates.
5. Once the workflow runs without you, move to the next task.
> “Don’t wait for the perfect stack. Build, document, then iterate.” — Grace Leung
Source: [Ryan Doser interview with Grace Leung](https://www.youtube.com/watch?v=HBKKMC5RpJw)