Thin Harness, Fat Skills: The New Way To Build Software
YouTube · 57lDpTwiW6g
Quick Read
Summary
Takeaways
- ❖Gary Tan, after a 13-year coding break, shipped hundreds of thousands of lines of code in months by leveraging AI agents, achieving 400x his previous output.
- ❖The "Thin Harness, Fat Skills" paradigm advocates for minimal, stable code (harness) to orchestrate powerful, markdown-based AI agents (skills).
- ❖"Token maxing" – providing LLMs with maximum context and allowing them to "boil the ocean" – is crucial for comprehensive research and complete code generation.
- ❖AI agents can achieve 80-90% test coverage, automating the least enjoyable part of coding and significantly improving software quality.
- ❖The current AI development environment is compared to the "Homebrew Computer Club" era, requiring builders to be "mechanics" for their powerful but sometimes brittle tools.
- ❖Embracing personal AI and learning to write effective prompts is essential for maintaining control over one's tools, rather than being controlled by corporate algorithms.
- ❖Founders should view token spend as "rent" – a critical investment for competitive advantage, not an expense to economize.
Insights
1400x Productivity Boost with AI Agents
Gary Tan, after a 13-year hiatus from coding, returned to building and achieved a 400x increase in logical lines of code output compared to his previous coding peak. This was accomplished by directing 15 AI agents simultaneously, primarily using Claude Code and OpenClaw, to generate and test code.
It was 13 years of not coding and then suddenly boom, I'm doing about 400x the amount of work... I was directing you know 15 agents at a time to do so.
2The "Thin Harness, Fat Skills" Architectural Philosophy
The core idea is to build a "thin harness" – a minimal, robust code loop that handles user input and orchestrates LLM actions – and "fat skills" – complex, context-rich instructions written in markdown that leverage the LLM's latent space for flexible, intelligent execution. This separates deterministic actions (code) from nuanced, knowledge-based tasks (markdown/LLM).
A harness is the core loop that takes the user input gives it to the LLM runs what the LLM does... What we should spending all our time doing is thinking about what markdown should there be?
3"Token Maxing" for Comprehensive AI-Driven Work
To achieve high-quality, complete results with AI agents, it's essential to "token max" – provide the LLM with the most extensive possible context, allowing it to "boil the ocean" in terms of research or code generation. This means feeding it dozens of sources, cross-referencing information, and accepting higher token costs for superior output.
What if you absolutely boiled the ocean... if there is incremental work that makes something more complete more awesome... you should token max.
4AI Automates Comprehensive Testing and QA
AI agents can achieve 80-90% test coverage (unit, integration, end-to-end) for generated code, automating the typically tedious and time-consuming aspects of software development. This significantly improves code quality and allows human developers to focus on higher-level tasks.
I can get to 100% test coverage... hitting 80 to 90% is usually the best practice... the machine doesn't care, it'll just do it.
5The "Ferrari with a Wrench" Analogy for Current AI Tools
Using advanced AI tools like OpenClaw is likened to driving a Ferrari: exhilarating and powerful, enabling unprecedented feats. However, like a Ferrari, these tools are currently brittle and will break down, requiring the user to be a "mechanic" and fix them manually. This highlights the need for deep technical skill alongside AI adoption.
Using OpenClaw these days is like driving a Ferrari... it's a Ferrari that will break down... you need to get out with your wrench and pop the hood and like fix it.
Bottom Line
The current AI development phase is akin to the "Homebrew Computer Club" era, where powerful but unpolished tools (like the Apple I) required users to be highly technical "kit car Ferrari" builders to get them running.
This implies that early adopters and those willing to get their hands dirty with current AI tools will gain a significant advantage, shaping the future of software development.
Develop tools, educational content, or communities that simplify the "mechanic" aspect of AI agent development, making these powerful tools more accessible to a broader audience without sacrificing control.
The defining question for the personal AI revolution is whether individuals will control their own AI tools and data, or if corporate entities will dictate the experience, similar to social media feeds.
This is a call to action for users to learn prompt engineering and take ownership of their AI agents to prevent a future where personal AI is centrally controlled and optimized for external business models.
Build open-source, privacy-preserving personal AI frameworks and educational platforms that empower individuals to customize and control their AI agents, fostering a decentralized AI ecosystem.
Opportunities
AI-Powered Investigative Journalism Platform
A platform that uses agentic retrieval (Perplexity, X, Grok APIs) to "boil the ocean" on any topic, ingesting vast amounts of internet data, cross-referencing sources, and generating detailed, fully sourced long-form articles and reports. It acts as an investigative journalist, not just a publishing tool.
AI Agent Workflow Orchestration and Skill Repository
A system (like GStack) that allows users to define and chain AI "skills" (e.g., CEO skill, Designer skill, QA skill) using markdown, enabling complex, multi-agent workflows for software development, product planning, or other knowledge work. It would manage agent interactions, context passing, and provide a human-in-the-loop interface.
AI-Driven Automated QA and End-to-End Testing Service
A service that leverages AI agents (e.g., wrapping Playwright) to automatically perform comprehensive QA, including end-to-end, integration, and unit testing, for new features or entire codebases. It would interpret context, simulate user flows, and report bugs, drastically reducing manual testing effort.
Key Concepts
Thin Harness, Fat Skills
This architectural principle suggests building minimal, stable code "harnesses" that handle core loops and deterministic actions, while delegating complex, context-dependent, and evolving tasks to "fat skills" defined in markdown for AI agents. The harness provides structure, and the skills provide flexible, intelligent execution.
10x Check / 10-Star Experience
A product and design exercise, inspired by Brian Chesky, that pushes beyond conventional improvements to imagine a 10x more ambitious solution or a "10-star" user experience. Applied to AI, it encourages thinking about how AI can deliver disproportionately higher value (10x value for 2x effort) by leveraging its latent space capabilities.
Lessons
- Embrace "token maxing": Don't economize on token spend when using LLMs for critical tasks; invest in providing maximum context to achieve more complete and higher-quality outputs.
- Adopt the "Thin Harness, Fat Skills" architecture: Design your AI-powered applications with minimal, stable code harnesses and powerful, markdown-defined AI skills for flexibility and scalability.
- Integrate AI for comprehensive testing: Leverage AI agents to automate unit, integration, and end-to-end testing, aiming for 80-90% coverage to ensure robustness and free up human developers.
- Cultivate "mechanic" skills: Be prepared to debug and fix AI agent failures, as current tools are powerful but brittle; this hands-on approach is crucial for early adopters.
- Learn prompt engineering: Actively develop your ability to write effective prompts and define AI agent behaviors to maintain control over your personal AI tools and avoid reliance on opaque corporate algorithms.
GStack AI Agent Development Workflow
Define Core Idea (Office Hours Skill): Start with a high-level product idea, clarifying "who is it for, what does it do, and what's the impact?"
Strategic Planning (CEO Skill): Apply the "10x Check" and "10-Star Experience" mental models to envision the platonic ideal and most ambitious version of the feature, aiming for 10x value for 2x effort.
Architectural Design (ASCII Diagrams): Before coding, instruct the AI to generate ASCII diagrams of data flows, state machines, dependency graphs, and user flows to load context and ensure completeness.
Developer Experience & UI Review (Designer/Developer Skills): If applicable, run design and developer experience reviews to ensure usability and ease of integration.
Code Generation & Testing (Plange/Codex Skills): Use AI agents (like Claude Code) to generate code, prioritizing 80-90% test coverage (unit, integration, end-to-end) to prevent "slop."
Expert Review & Debugging (Codex/Claude Switch): For complex problems or persistent bugs, switch between different AI agents (e.g., Codex for "200 IQ nearly nonverbal CTO" problems, Claude for "ADHD CEO" oversight) to find solutions.
Human-in-the-Loop QA (Browse/QA Skill): Manually test UI or data mutations using an AI-powered blackbox browser (like Browse) while providing human feedback, as the human operator's understanding remains irreplaceable.
Notable Moments
Gary Tan's return to coding after 13 years, shipping hundreds of thousands of lines of code while running YC full-time.
This demonstrates a paradigm shift in developer productivity enabled by AI agents, challenging the notion that such output is impossible for a single individual, especially one with a demanding executive role.
The realization that AI can automate comprehensive software testing (80-90% coverage), a task typically disliked by developers.
This highlights AI's potential to improve software quality and free human developers from tedious work, allowing them to focus on more creative or strategic tasks.
The comparison of current AI development to the "Homebrew Computer Club" era and "kit car Ferrari" phase.
This frames the present as a foundational moment for personal AI, emphasizing that early adopters who are willing to grapple with imperfect tools will be at the forefront of a new technological revolution.
Quotes
"Will you have control over your own tools or will your tools have control over you?"
"Using OpenClaw these days is like driving a Ferrari and it's like exhilarating. It's insane... But then it's also like a Ferrari and that you better be a mechanic. like it's a Ferrari that will break down on the side of the road... and you need to get out with your wrench and pop the hood and like fix it."
"If you token max like that's actually the coolest thing you can do now and it's not just in you know generating articles it's not you know it's clearly in uh writing code right I think now it's it's going to permeate every part of society like every thing that we would call knowledge work could be token maxed."
"I just want the machine to do the stuff that I don't want to do."
"You could have infinite time by borrowing the time from the machines."
Q&A
Recent Questions
Related Episodes
DAPHNIQUE SPRINGS | ENGLISH MAJORS | SEASON 3 | EP 10
"Comedian Daphnique Springs joins the hosts for a freewheeling discussion on everything from her childhood fish-selling hustle and the 'Florida Man' phenomenon to deep-sea conspiracies, the future of AI, and parenting advice for children with ADHD."
Code Health Guardian
"This talk introduces a comprehensive model for understanding and managing code complexity, arguing for its objective nature and the critical role of human understanding in the AI era to maintain software health."
Wes Watson Student Arrested as a Firefighter
"A former firefighter and paramedic recounts his harrowing descent into drug addiction, multiple brushes with the law, and ultimate recovery, culminating in a successful business partnership and a transformed life."
The Joe Budden Podcast Episode 926 | Let’s See That Snapper
"The hosts dissect the collapse of Spirit Airlines, the predatory nature of music contracts, and the weaponization of private sexual preferences in public disputes, all while celebrating the Knicks' playoff dominance."