paul-graham-perspective

Paul Graham Perspective turns the assistant into a thinking advisor that analyzes startups, writing, products, and life choices through Paul Graham's mental models and voice. According to the skill, it is distilled from research across 200+ essays, roughly a dozen podcasts and interviews, Twitter/X analysis, seven core critics, and a full life timeline, yielding five core mental models, eight decision heuristics, and an expression style. It activates on cues such as using PG's perspective, what Paul Graham would think, PG mode, or paul graham perspective, and even looser phrasings that ask to think from PG's angle.

Once active, it responds in the first person as Paul Graham rather than describing him in the third person, with a defined exit trigger that drops the persona on request. Its distinguishing design is an agentic protocol that resists answering from memory alone. Incoming questions are first classified as fact-dependent, pure-framework, or mixed. Fact-dependent questions (specific companies, people, products, or post-2024 events) require real research via web search before answering, evaluating whether founders are makers or managers, whether a market is small but fast-growing, whether users love rather than merely like a product, and whether growth is organic. Pure-framework questions skip straight to reasoning from the mental models, and checkpoints between steps guard against faking knowledge or dumping a research brief instead of a judgment.

The mental models it draws on include Writing equals Thinking, Taste as a Cognitive Instrument, Iterative Discovery, and Superlinear Returns, each stated with supporting evidence, applications, and its own limits. A scene-to-model lookup routes questions about startups, writing, careers, evaluating people, or time management to the most relevant model and heuristic. Extensive failure-mode and anti-pattern lists enforce the voice: lead with a headline judgment, allow honest hedging like admitting a topic has not been thought through enough, avoid academic jargon and listicles, and never claim to know a person or company that has not actually been researched.