AI Revisited - part 2

What It Covers

Jason Fried, CEO of 37signals, shares how he uses AI tools Claude and ChatGPT daily for writing, prototyping, and generating test data, while explaining the company's deliberate strategy of waiting before embedding native AI features into Basecamp, Hey, and Fizzy products.

Key Questions Answered

• •AI as language alignment tool: Feed AI a corpus of real customer language — such as ~1,000 testimonials — then ask it to audit your copy for terminology mismatches. Fried discovered customers say "organized" where he wrote "stay on top of things." This closes the gap between internal product vocabulary and how buyers actually describe value.
• •Rapid prototyping without developer dependency: Use Claude to build functional local prototypes before assigning engineer time. Fried built a working "recent contacts" sidebar feature in Basecamp himself to validate the concept first. This avoids pulling teammates off priorities, especially critical when leadership requests carry unintended organizational weight regardless of stated urgency.
• •Synthetic test data generation via prompt: Instead of manually logging in as multiple users to populate realistic UI states, prompt Claude to generate weeks of multi-user chat history — including file attachments, varied message lengths, and emoji reactions — directly into a local database. This produces usable design context in seconds rather than hours.
• •Agent-first product strategy over native AI features: Rather than building custom AI features that may become obsolete, 37signals is developing CLI access and cleaner data interfaces so third-party agents like OpenAI or Anthropic offerings can operate as standard users inside Basecamp. Fried cites OpenClaw as an early signal that always-on personal agents will soon be mainstream.
• •Human support as competitive advantage at small scale: With a 62-person company where roughly one-third are engineers and designers, 37signals treats long-tenured support staff — some with 15 years on the team — as a structural differentiator. Routing all support through AI bots before humans is explicitly rejected as a model, with direct email to humans kept as the primary channel.