Source: raw/Codex_Claude_Code_-_Build_Your_AI_Marketing_System_that_Works_on_Any_AI.md (https://www.youtube.com/watch?v=cAeKDWa77xI), YouTube auto-captions fetched 2026-10-30. Creator: Grace Leung (youtube.com/@graceleungyl) Platform: YouTube Duration: 16:54 Published: 2026-07-17 Skill level: Intermediate (local folders, MCP config, two desktop agents)

Grace Leung argues that the gap between frontier models is getting small, so the asset worth building is the harness around the model, kept in a form you can carry between vendors. She builds a marketing agent workspace in Claude Code, opens it in OpenAI’s Codex, adds a skill and an agent there, and syncs it back to Claude Code with a migration skill she wrote. The video is sponsored by HubSpot, which co-produced the skill pack it promotes.

Key Takeaways

  • Model vs harness. “The model determines its intelligence in doing marketing tasks, but the harness is what makes it your own marketing agent.” Swap Claude for a ChatGPT model and the marketing ability changes, but the brand voice, data access and workflows stay yours.
  • What makes an agent is the loop: take an action, check the result, keep going until the job is done. She attributes the framing to NVIDIA’s Jensen Huang.
  • Cloud agents vs local agents. Cloud agents run on the vendor’s servers and don’t automatically see your computer; she lists Claude Managed Agents, Google Gemini Spark, Google Enterprise Agent Platform, Perplexity Computer and Manus. Local agents read your files directly; she lists Claude Code, Cowork, Codex, Hermes and OpenClaw. The video covers local agents.
  • Five components of the workspace:
    1. Knowledge and context, including memory
    2. Instructions or configuration files
    3. Agent skills
    4. Specialized subagents
    5. Tools (MCP connections to analytics, CRM and so on)
  • Three portability buckets:
BucketComponentsWhat moving takes
1. Fully portableContext and knowledgeNothing; they are plain files
2. Small translationInstructions, skills, subagentsSame content, different file names and default folders
3. Not directly portableToolsReconnect on the new platform
  • Inherit by default, add only what’s different. One shared context folder and shared skills, agents and instructions at the top. Below that, one folder per channel for a marketing team (paid, content, email, SEO), per type of work for a functional team, or per client for a consultant. A client with its own needs gets its own context and skills in its own folder.
  • Codex imports a Claude Code project for you. Opening the folder in Codex for the first time, it detected the Claude configuration and offered a one-click import. It created a .agents folder, a .codex folder, a config file and AGENTS.md, and the skills panel showed every skill migrated. Check MCP connections afterwards and reconnect any that dropped.
  • New work in Codex lands in Codex’s folders. A landing-page-audit skill installed into .agents; a new growth-specialist agent was saved under the Codex folder. Moving back therefore needs a conversion.
  • Codex’s browser control impressed her. “From my testing, Codex is much faster in controlling the browser and is much more accurate as well.” Her growth agent opened the homepage, took screenshots at several resolutions and wrote a Word-format audit with fixes and an action list.
  • Sync back with a migration skill, after a backup. Her free migration skill, installed at user level in both apps, scans the project, proposes a migration plan, asks whether to back up and then converts. “I always recommend you to make a backup for agent workspace first.” Back in Claude Code, the Codex-made growth agent read Codex’s audit and redesigned the homepage with the brand-design skill.
  • The skill doesn’t sync agent memory. Her suggestion: point both platforms’ memory files at “one single source of truth” that both can read.
  • Her conclusion: “Don’t care too much about the model’s intelligence. Instead, set up a better foundations for your agent system so it can compound over time.”

Try It

  1. Sort your current setup into the five components. List which ones exist only inside one vendor’s app (connectors, memory) — that list is your lock-in.
  2. Move brand context into plain files in one shared folder if it lives in chat projects today.
  3. Keep skills in the open SKILL.md format and agents as Markdown role files, so moving them only means changing folders.
  4. Open the same folder in a second agent app and run one real task. Note what broke; it is usually a tool connection.
  5. Before converting folders in either direction, commit or copy the workspace.

Implementation

Tool/Service: Claude Code (shown in VS Code), OpenAI Codex desktop app, Google Search Console through MCP, her migration skill, the HubSpot “Marketer’s Agent Skill Pack”.

Setup (the Claude Code side as shown):

marketing/
  CLAUDE.md          # project instructions
  .mcp.json          # MCP connections (e.g. Google Search Console)
  core-context/      # brand, product offerings, audience
  seo/  content/ ... # working folders
  .claude/
    skills/          # drag unzipped skill folders here
    agents/          # e.g. data-analyst.md (memory enabled)

After Codex import (as she describes it): a .agents/ folder (where Codex keeps skills), a .codex/ folder (where her new agent was saved), a config file, and AGENTS.md.

Skill pack contents (sponsor resource): five skills for content research, copywriting, strategy proposal, data visualisation and landing-page auditing. Each checks the brand context first.

Cost: Not stated. The skill pack and migration skill are free downloads linked from the video.

Integration notes:

  • Instructions now need less translation. Since v2.1.277 (2026-09-18), Claude Code reads AGENTS.md when a project has no CLAUDE.md (see AGENTS.md). The video predates this. A project kept only in AGENTS.md now loads in both apps without a converted copy; a project with both files still needs them kept in step.
  • Skills are a shared format. The Agent Skills standard published at agentskills.io is adopted by OpenAI among others (see Skills Ecosystem), which is why skills land in bucket 2, not 3.
  • Nate Herk’s manual route. Ask the agent “help me figure out what files to create to make this compatible with Codex”; usually it’s renaming claude.md to agents.md plus minor tweaks (see Master 97% of Codex in One Hour).
  • A different way to combine the two. Codex inside Claude Code keeps one workspace in Claude Code and calls Codex as an executor through OpenAI’s plugin, instead of switching apps.

Open Questions

  • The Codex auto-import of Claude configuration is shown on screen only. No first-party OpenAI source for it is in the wiki, and the folder names may differ between Codex versions.
  • She doesn’t show the migration skill’s conversion rules (for example, how agent frontmatter fields map between the two apps).
  • Several captioned names are uncertain: “Clockr” is probably Claude, and “Clerkbot”, listed beside Codex as another target, is unidentified.
  • The browser-speed comparison is one person’s test, with no numbers.
  • Google Gemini Spark and Google Enterprise Agent Platform aren’t otherwise covered in this wiki; they appear here only as names in her list.