Thin Harness, Fat Skills: The New Way To Build Software
This episode explores the transformative shift in software development, where AI agents empower solo developers to achieve unprecedented productivity. Featuring YC President Gary Tan's return to coding after 13 years, it delves into new workflows, the 'tokenmaxxing' philosophy, and the rise of personal, user-controlled AI.
A new era of software development has arrived, characterized by a symbiotic relationship between human engineers and AI agents. This shift enables a single person to achieve what previously required entire teams and millions of dollars in funding. This transformation is not just about productivity tools; it’s about a fundamental change in how software is conceived, built, and maintained.
The Return to Building: A 400x Leap in Productivity
After a 13-year hiatus from coding, YC President Gary Tan returned to building, leveraging AI tools like Claude Code. He rebuilt his first startup, Posterous, in just five days with a $200 budget. This feat, which originally took a team of seven, a year and a half, and $4 million, highlights a staggering increase in efficiency. This new version was not just a simple blog platform but a sophisticated system with full RAG (Retrieval-Augmented Generation) capabilities, capable of performing deep, journalistic research on any topic by ingesting and synthesizing vast amounts of information from the internet.
This experience led to the observation of a roughly 400x increase in output. This isn’t just about writing more lines of code; it’s about directing agents to produce tested, production-ready logical lines of code at a scale previously unimaginable for an individual.
“Tokenmaxxing”: The Philosophy of Boiling the Ocean
A core principle behind this new paradigm is “tokenmaxxing”—the idea of using a large volume of tokens to achieve the most complete and high-quality result possible. Instead of taking shortcuts, the AI is prompted to “boil the ocean.”
- Deep Research: For a given topic, an agent can be instructed to ingest not one or two sources, but twenty or more, cross-referencing them to understand different perspectives and create a comprehensive, nuanced output.
- Complete Context: This applies to coding as well. By feeding the AI extensive context, including architectural plans, data flow diagrams, and user flows, it produces more robust and complete code, reducing bugs and oversights.
This approach transforms knowledge work by enabling a level of thoroughness that would be prohibitively time-consuming for a human. It’s about leveraging the machine’s capacity to process vast information to make better decisions and create superior products.
A New Workflow: The Rise of the Agentic Engineer
This new way of working has given rise to novel tools and workflows, exemplified by the creation of GStack, an open-source project that emerged from personal productivity hacks.
- Structured Planning: Before writing any code, the process begins with high-level planning “skills.” Prompts like the “CEO plan” (inspired by Brian Chesky’s “10-star experience” exercise) force the model to think about the platonic ideal of a feature, pushing for a 10x value increase for only 2x the effort.
- Visualizing Before Building: A key discovery was that asking the AI to first generate ASCII diagrams of data flows, state machines, and dependency graphs dramatically improved the quality of the final code. This pre-computation step forces the model to load the full context and structure of the problem into its latent space.
- Automated Testing and QA: AI agents excel at the tedious but critical task of writing tests. It’s now feasible for a solo developer to achieve 80-90% test coverage, a best practice often neglected due to time constraints. This extends to QA, where agents can wrap tools like Microsoft’s Playwright to perform automated, black-box browser testing of new features.
- Multi-Agent Collaboration: Different models have different strengths. The workflow involves using multiple agents in concert:
- Claude: Acts as the “ADHD CEO,” ideal for brainstorming, high-level planning, and creative tasks.
- Codex: Serves as the “200 IQ, nearly nonverbal CTO,” brought in to perform deep, rigorous analysis of the codebase to find bugs and logical flaws that other models might miss.
Thin Harness, Fat Skills: A New Architectural Paradigm
The philosophy for building these agentic systems is “Thin Harness, Fat Skills.”
- Thin Harness: The core application code that loops and interacts with the LLM should be as simple, minimal, and deterministic as possible. There is no need to reinvent the wheel.
- Fat Skills: The complexity, nuance, and intelligence of the system should reside in the prompts, which are written in Markdown. These “skills” are like well-documented checklists or instructions that guide the LLM’s reasoning process. In this sense, Markdown becomes a form of high-level, human-readable code that is “compiled” by the LLM.
The art of agentic engineering lies in discerning what logic belongs in brittle, deterministic code versus what belongs in the flexible, context-aware latent space of the model.
The “Kit Car Ferrari” Era and the Future of Personal AI
Working with today’s AI agents is like driving a Ferrari: exhilarating and incredibly powerful, but also prone to breaking down. The user must be a “mechanic,” ready to pop the hood and fix problems. We are in the “Homebrew Computer Club” era of AI—a time of immense potential for those willing to get their hands dirty.
This leads to a defining question for the future: Will we have personal, user-controlled AI, or will it be dominated by large, corporate-controlled systems with opaque algorithms? The most powerful path forward is personal AI, where individuals can write their own prompts and tailor systems to their unique needs and values. By embracing this, we are not just building software; we are buying back time, leveraging machine consciousness to become “time billionaires” and unlock human potential on an unprecedented scale.