Why I stick to Claude for work (most of the time)
The creator demonstrates why Claude is their preferred AI for work by comparing it with ChatGPT across multiple tests, showing that Claude better adheres to custom instructions and agentic workflows defined in local folder structures. The key advantage lies in using organized, LLM-agnostic folder systems with agents.md files rather than relying on auto-memory features.
Summary
The video compares ChatGPT and Claude for professional work, specifically testing how well each AI follows custom instructions and context from local folders. The creator first disables auto-memory and random memory generation in both applications to ensure a fair comparison, then runs identical prompts through both LLMs using their desktop applications.
In the first test without folder context, both GPT-4 Astra and Claude 3.5 produce generic LinkedIn posts about calendar task management with similar quality and speed. However, when the creator points both AI systems to a folder containing their personal productivity system (myICOR methodology with agents, guidelines, and workflows), the outputs diverge significantly.
Claude demonstrates superior orchestration by immediately recognizing the agents.md file structure, spinning up a sub-agent named Sage (the LinkedIn specialist), reading the sub-agent's instructions, and producing content that deeply reflects the creator's actual teaching about calendar vs. task list management. ChatGPT/Codex also attempts orchestration but produces more generic output that doesn't incorporate the specific philosophy and guidelines stored in the folder.
The creator then repeats the test in the terminal using CLI versions of both tools (OpenAI Codex and Claude) to eliminate any desktop application interference. Claude again demonstrates better adherence to custom instructions, with the creator noting that Claude is "too verbose" at times but consistently "sticks to the rules much more" compared to other LLMs.
The video emphasizes that the real power comes from understanding agentic workflows: a root orchestrator agent (Larry) delegates work to specialized sub-agents (Sage for LinkedIn, Charter for design, Buzz for social media), each with their own agent.md file. These agents reference shared guidelines, SOPs (standard operating procedures), and work streams stored in organized subfolders. This structure is token-efficient because it loads only relevant information for specific tasks rather than keeping everything in one massive prompt.
The creator demonstrates this folder architecture in detail, showing how it's designed to be LLM-agnostic. The claude.md file points to agents.md (now the standard), which allows ChatGPT, Claude, and Gemini to all reference the same folder structure. This approach provides independence from any single LLM—if Claude underperforms in the future, the creator can quickly switch to another model while maintaining the same workflow system.
About this episode
Memory off in ChatGPT and in Claude, the same folder, the same prompt: only one of them followed the rules inside it. 👉 The folder in this video is free inside myICOR: https://www.myicor.com/?utm_source=youtube&utm_medium=video&utm_campaign=LEeuIPaLjl4&utm_content=description WATCH THIS NEXT Why I switched Claude's memory off in the first place, and what the folder does instead: https://youtu.be/p79T9P2WmTI ABOUT THIS VIDEO Both desktop apps, memory and custom instructions switched off and deleted to be sure. Nothing carries over from earlier chats. Whatever the AI knows about me has to come from the folder. Three tests, the same prompt every time: a LinkedIn post about the pros and cons of managing your tasks on a calendar. First with no folder at all. Then with my folder open as the project, in the ChatGPT desktop app and in Claude Code inside the Claude desktop app. Then once more in the terminal, Codex on one side and Claude Code on the other. With the folder open, Claude reads the agents.md file, takes the orchestrator role, hands the post to the LinkedIn specialist inside the folder and comes back with a post that says what I actually teach: the calendar holds constraints, the task list holds freedom. ChatGPT on the same folder gives a generic post that could be anyone's. In the terminal the same thing happens again, and Claude even names its own mistake when I push back. Then the part that matters: this is not the model. Auto memory keeps what you tell it and that is fine for simple work. A local folder with contracts, an agent index, guidelines and procedures is what makes the same quality repeat every day, and only the parts that are needed get loaded, which saves tokens. Claude now reads agents.md, so the same folder works with Codex, Gemini and Claude, and I can switch the moment one of them falls short. And it does fall short. For a slide deck, GPT 6 Astra gave me the better visuals, faster. Fable overthinks and takes longer. That is the "most of the time". The folder is what stays. CHAPTERS 00:00 Memory off in ChatGPT and in Claude 01:16 Test 1: same prompt, no folder 02:42 Test 2: my folder as the context 05:36 Claude reads agents.md and delegates 07:11 The two LinkedIn posts side by side 09:46 Test 3: Codex vs Claude Code in the terminal 10:56 Orchestrator and sub-agents in the terminal 13:15 Why Claude sticks to the rules 14:11 Auto memory vs a local folder 14:53 Inside the folder: agents.md and the orchestrator 16:20 Agent index, guidelines, SOPs, workstreams 18:02 Why this saves tokens 19:22 Why I stick to Claude 20:33 Where ChatGPT still wins 21:32 One folder, any LLM: the free scaffold and the course WHAT I AM USING The Claude desktop app (Cowork and Claude Code) and the ChatGPT desktop app side by side, then Codex and Claude Code in the terminal. The top models on both sides, GPT 6 Astra on medium effort and Fable 5.1. Memory and custom instructions switched off in both apps so the memory lives in the folder instead. The folder is the free ICOR for Life scaffold with the AI team inside it. You can download the folder for free here: https://www.myicor.com/?utm_source=youtube&utm_medium=video&utm_campaign=LEeuIPaLjl4&utm_content=description The video I mention at the start, where Claude Cowork and Claude Code get the same folder: https://youtu.be/PxpVh7Dsw1M ABOUT Tom helps professionals build AI productivity systems that work, using the ICOR Methodology, courses, coaching, and the myICOR community. LinkedIn: https://www.linkedin.com/in/tomsolid/ X: https://x.com/TomSolidPM Podcast: https://www.youtube.com/@productivitylikeapro All third-party marks and logos shown are the property of their respective owners and are used for identification in editorial commentary only. No partnership or sponsorship is implied. #Claude #ChatGPT #ClaudeCode #ICOR #myPKA
Key Insights
- Claude consistently adheres to custom instructions stored in agents.md files better than ChatGPT, even when both models are pointed to identical folder structures, demonstrated by Claude's deeper incorporation of the creator's specific ICOR methodology in generated content.
- Disabling auto-memory in AI applications and instead building memory locally within folder structures allows users to control exactly what information the AI retains and when, rather than having the AI randomly decide what to keep and discard.
- The agents.md file has become the new standard for both Claude and ChatGPT (replacing claude.md), making folder structures LLM-agnostic and allowing users to switch between different language models without restructuring their entire system.
- Agentic workflows with specialized sub-agents are more token-efficient than loading all instructions into a single agent.md file because the orchestrator only loads relevant guidelines and SOPs for specific tasks, preventing system degradation as complexity increases.
- Claude demonstrates superior transparency and usability in multi-agent orchestration by allowing visual switching between sub-agents in the terminal interface, whereas ChatGPT's orchestration is less transparent to the user.
Topics
Transcript
[0:00] In this video, I want to talk about Personalized Memory of your AI agents and the LLM that you are using. And here you see on the right side, it's the desktop application of Claude, and on the left, we have the desktop application of ChatGPT. And here I'm in the settings and the Codex instructions are empty. this means there are no instructions that it will pick up the moment I start a chat. And also the Codex memory is switched off, and I even clicked here on delete to ensure it's really empty. And here on Claude, you see the same thing. [0:30] I switched off both. Right. There's no search and reference chats. There is no…
Full transcript available for MurmurCast members
Sign Up to AccessMore from ICOR with Tom | AI Productivity
You are using the wrong Claude for work! (Here is proof)
A video demonstrating that Claude Code is significantly more powerful than Claude Cowork for professional knowledge workers, showing how Claude Code better understands folder context, enforces safety rules, orchestrates sub-agents with different models, and creates persistent file outputs rather than temporary artifacts.
Obsidian is the tool I teach in. Keep the one you have.
The creator addresses criticism about switching to Obsidian after previously promoting other tools, clarifying that his tool-agnostic ICOR methodology remains unchanged and that Obsidian is simply an interface layer for his folder-based system. He emphasizes that the underlying folder structure, AI automation, and productivity principles are identical regardless of which tool visualizes the content.
Claude sucks. And here is why.
This video explores why Claude and other AI models produce inconsistent results, examining how model selection (Haiku vs Sonnet vs Opus vs Claude 3.5 Fable), effort levels, prompt clarity, and system context dramatically affect output quality. The speaker demonstrates these differences through a Venn diagram creation task and advocates for using organized folder systems with documented SOPs and code-based solutions to ensure consistent AI performance.
Claude can now prompt its own SESSIONS. I built the boss.
Claude now has the ability to interact with multiple sessions, allowing a streamlined orchestration of tasks through an agent called Larry. This integration can simplify task management and enhance productivity by enabling a single main session to oversee various sub-sessions or tasks.
I switched off Claude's memory. My folder does it better.
Tom explains why he disabled Claude's native memory feature and instead built a local folder-based system (myPKM) to maintain persistent context with AI agents. He shares an audit revealing systemic issues like contradictory documentation, forgotten rules, and parallel session conflicts, then outlines how to systematically fix these problems using the ICOR productivity methodology.