# 37signals AI Recommendations Key insights from [[David Heinemeier Hansson (DHH)]] discussion with Jeremy about the role of AI at 37signals, shared in the post "The AI conversations" (February 2026). ## Outcome-Based Instructions Switch your mode of instruction from being task-based ("implement this like this") to becoming outcome-based ("fix this", "build this"). Agents are not just ways of accelerating your designs, but also increasingly capable of coming up with great designs on their own. ## Run Dangerously Graduate to "run dangerously". You're not getting the full experience of agents if you're having to babysit every little step they do with permissions. The way to protect your system is by ensuring that your machine is disposable, so even in the case of catastrophe, you can be back up and running in a few minutes. DHH uses the alias `cc="claude --dangerously-skip-permissions"` for this purpose. ## Avoid Upfront Planning Overly detailed, upfront planning is a bit of an antipattern with agents, just like it was with humans. You don't know what you want until you see something running. So don't be afraid of kicking off exploratory expeditions, then asking AI to refine it. This even goes for code style. Agents take revisions and direction very well. Then ask it to commit those style guidelines to the AGENTS.md file. ## Embrace Git Worktrees Learn to embrace Git worktrees! Waiting on agents is the new "waiting on the compiler", but you can use the time by managing multiple expeditions concurrently. Just be careful you don't start so many spinning disks that you fry your brain with multitasking! ## Continuous Agent Operation Tools like [[OpenClaw]] point to a new future where agents just run continuously. This points to how we'd want to interact with agents in project management: Assigning high-level work, then having them report back. ## Right Level of Granularity What is project management actually for? At its best, it's about sharing progress, outstanding tasks, and offering opportunity to chime in for humans at a level of granularity that people actually care about. It doesn't make sense to have agents use project management tools for every subtask, just like we don't do that as humans. We need to hit a level of granularity that serves the original mission statement: Project Management Is Communication. ## The Bespoke Future The future hints at software returning to the realm of the bespoke. Asking your agent to make software on your behalf becomes as easy/good as buying it off the rack. Which in turn points to a future where *everything* becomes bespoke. Your robot becomes your personal tailor, chef, and tutor. ## Disruption Theory in Action AI is so useful that hundreds of thousands of people are willing to run AI agents with direct access to their emails, passwords, and more, despite the known security pitfalls. It's classic disruption theory: Early adopters get excited by what others dismiss as toys/dangerous/slow processes, but the fundamentals are so strong that eventually the toys become tools, the danger gets mitigated, and the slow processes are sped up. And then the world is changed. ## Conclusion We're still in the early days of this, but it's definitively clear that, like the internet, we're not going back to the old world. Despite all the incessant, over-the-top hype that's frequently nauseating to watch, you shouldn't turn away. You should tune in, turn it on, and integrate it into your work. ## Source - [[David Heinemeier Hansson (DHH)]], "The AI conversations", 37signals Basecamp, February 3, 2026: https://x.com/dhh/status/2018631575337095389 ## Related - [[AI Agents]] - [[Claude Code]] - [[OpenClaw]] - [[AGENTS.md (File Convention)]] - [[Disruption Theory]]