Thin Harness, Fat Skills: Garry Tan’s AI Engineering Secrets for 1000× Efficiency

Why Your "Fat Harness" Is Killing Your AI

Garry Tan argues that most current agent frameworks suffer from two sins: an overly bulky harness layer and fragile, low‑level skills, which consume half the context window and slow reasoning.

Five Core Principles: From Prompt Writing to Programming AI

He identifies five architectural patches that embody the "Thin Harness, Fat Skills" philosophy.

Principle 1: Skill Files – Write Method Calls in Markdown

Excellent skill files behave like function calls: they accept parameters and produce varied capabilities. For example, the /investigate skill consists of seven steps (dataset scoping, timeline building, annotation, synthesis, debate, citation) and can turn a 2.1‑million‑email corpus into a medical‑research analyst or a shell‑company investigation into a legal analyst.

“This is not prompt engineering. It’s software design. You use Markdown as a programming language and human judgment as the runtime.”

Principle 2: Resolvers – AI’s Smart Routing Table

Resolvers decide when and which context to load. Tan once wrote a 20 000‑line CLAUDE.md file that degraded model attention; he split it into a 200‑line pointer file plus a resolver that reads docs/EVALS.md to run evaluation suites and roll back if accuracy drops more than 2%.

“When a developer changes code without a resolver, the change is submitted directly. With a resolver, the model first reads the evaluation file and conditionally rolls back.”

Principle 3: Latent Space vs. Deterministic Tasks

AI excels in latent‑space tasks (understanding, reasoning, pattern recognition) but fails on deterministic, combinatorial problems. For instance, an LLM can seat eight people intelligently but will hallucinate a wrong arrangement for 800 seats, which should be delegated to traditional algorithms.

Principle 4: Propensity Analysis – Giving AI Judgment

Propensity analysis transforms AI from a data retriever into an analyst. Using YC’s internal AI that monitors 6 000 founders, the system discovered a “FinOps” project that claimed to build a Datadog for agents but was actually writing billing modules for 80% of its code.

“The founder claimed to build a Datadog for AI agents; in reality, 80% of the code was a billing module.”

Principle 5: Skills as Permanent Upgrades

Every skill file becomes a lasting upgrade: once a task is codified in a skill, it persists and automatically benefits from future, more powerful models.

“You must never perform a one‑off task. If a task may be repeated, prototype it on 3‑10 samples, get approval, then solidify it into a skill file or scheduled job. If you need to ask the same thing twice, you have failed.”

Conclusion

When large‑model capabilities converge, the real moat lies not in API access but in the efficient, self‑evolving “external brain” built with a thin harness and rich, reusable skill assets.