Thin Harness, Fat Skills: The New Way To Build Software

Key Concepts

• Token Maxing: The strategy of maximizing the use of LLM tokens to achieve higher quality, more complete, and more researched outputs, effectively "boiling the ocean" of available data.
• Agentic Engineering: Building software systems where AI agents perform complex tasks (research, coding, testing, QA) with human oversight.
• Vibe Coding: A colloquial term for building software by interacting with AI agents in natural language rather than writing traditional code manually.
• GStack: A custom framework/skill repository developed by Gary Tan to automate repetitive development tasks, including architecture reviews, testing, and QA.
• Harness: The core loop that manages user input, LLM tool calls, and execution, acting as the interface between the human and the AI.
• Latent Space: The internal representation of knowledge within an LLM that allows it to understand context, user intent, and complex relationships.

---

1. Main Topics and Key Points

The video details Gary Tan’s return to building software after a 13-year hiatus, during which he shipped hundreds of thousands of lines of code while serving as the CEO of Y Combinator.

• The "Ferrari" Analogy: Using advanced AI tools like Claude Code is described as driving a Ferrari—it is exhilarating and powerful, but requires the user to be a "mechanic" who can pop the hood and fix the system when it breaks.
• Productivity Shift: Tan reports a 400x increase in output compared to his previous coding experience. He argues that "lines of code" remains a valid metric for productivity when standardized, as it reflects the volume of work directed through AI agents.
• The Personal AI Revolution: Tan posits that we are in a "Homebrew Computer Club" moment for AI. He advocates for individuals to build their own personal AI stacks to maintain control over their tools, rather than relying on corporate-controlled, black-box algorithms.

2. Real-World Applications

• Gary’s List: A platform built to address educational inequality (specifically access to algebra in San Francisco public schools). It functions as an "agentic newsroom," using AI to ingest vast amounts of internet data, cross-reference sources, and produce high-quality, sourced investigative journalism.
• Post Haven: A blogging platform rebuilt by Tan in five days for $200, replacing a project that originally took months and thousands of dollars.

3. Methodologies and Frameworks

• The "Plan-Review" Workflow: Tan emphasizes that AI agents perform better when forced to create an "ASCII diagram" of data flows, user flows, and state machines before writing code.
• The CEO/Developer Skill Set: Tan uses specific "skills" (prompts) to guide his agents:
  • CEO Skill: Uses meta-prompting to simulate high-level product thinking (e.g., "What is the 10-star experience?").
  • CodeX: A "200 IQ, nearly non-verbal CTO" agent used to audit code for bugs and architectural flaws.
  • QA Automation: Using Microsoft’s Playwright wrapped in an agent to automate end-to-end testing, allowing the agent to "browse" the site and verify functionality.

4. Key Arguments

• Human-in-the-Loop: Tan argues that while AI can handle the "drudgery," the human must remain the "agentic engineer" who provides taste, design direction, and the "why" behind the project.
• Thin Harnesses vs. Fat Skills: He suggests that the "harness" (the code that runs the agent) should be kept thin and simple, while the "skills" (the markdown-based instructions and prompts) should be "fat" and detailed, containing the logic and context of the task.
• Token Maxing as Rent: Tan compares token spending to paying rent in San Francisco—it may seem expensive, but it is "more expensive not to live there." Investing in tokens is an investment in utility and speed.

5. Notable Quotes

• "Will you have control over your own tools or will your tools have control over you?"
• "You could have infinite time by borrowing the time from the machines."
• "If you have taste and you understand technology, you are particularly the people who would benefit the most from getting this [AI]."

6. Synthesis and Conclusion

The core takeaway is that we are entering a "Golden Age of Open Source" where the barrier to building complex, high-quality software has collapsed. By adopting an "agentic" mindset—where the human acts as a director of multiple specialized AI agents—individuals can achieve levels of productivity previously reserved for large engineering teams. The future of software development lies in "token maxing" and maintaining personal control over one's AI stack, ensuring that the machine serves the human's unique goals and values.