Y Combinator CEO Shares Secrets of Building Highly Personified AI Agents
Y Combinator CEO Garry Tan posted that the core of creating highly personified AI Agents is not a single file, but a combination of three files: SOUL.md, USER.md, and AGENTS.md.
SOUL.md defines the Agent's identity, voice, values, and output standards (such as mandatory brevity, humor, and avoiding template openings). USER.md deeply models user thinking, blind spots, and preferences in about 4000 words, while AGENTS.md sets operational rules, failure handling, and lookup links.
In terms of market mechanisms, founders and developers are accelerating the adoption of a constitutional multi-file prompting approach in pursuit of personalized Agents. Under event-driven conditions, resources are shifting from generic system prompts to deeply personalized Agent frameworks, benefiting open-source tools like GBrain and the prompt engineering ecosystem, while standard chatbot products face pressure.
Source: Public Information
ABAB AI Insight
Garry Tan previously open-sourced the GBrain project as the brain system for his actual AI Agents, having iterated SOUL.md multiple times to inject specific rules like 'brevity mandatory' and 'humor mandatory', and tested multi-agent collaboration in Claude Code and OpenClaw environments. This path continues his long-term behavior as YC CEO promoting founder AI toolchains.
On the capital path, Tan mobilizes community contributions and YC network resources through open-sourcing GStack and GBrain, quickly validating the effectiveness of the multi-file architecture in production environments. The motivation is to upgrade personal Agents from generalized chat tools to highly aligned 'smartest friend' simulators, helping founders reduce context-switching costs and amplify daily execution efficiency, while providing replicable templates for AI-native companies within the YC ecosystem to attract more early adopters.
Similar to the custom Agent frameworks promoted by institutions like Andreessen Horowitz, and the evolution of early Prompt Engineering from single system prompts to multi-layer memory systems, Garry Tan's current method shows that AI Agent development is transitioning from reliance on generic large models to personalized constitutional control. Essentially, this represents a technological substitution: the multi-file constitutional architecture replaces traditional prompt engineering with a structured separation of identity + user + operations, where the mechanism of specific and strongly opinionated SOUL.md significantly reduces model hallucinations and enhances output consistency, forcing developer resources to shift from repeatedly debugging generic prompts to building deep context once, achieving a structural reconstruction from shallow interactions to persistent personalized Agents.