My Claude and Codex limits were gone in 2-3 days. Codex kept forgetting what the other sessions had done. For three weeks I couldn't ship one feature, and I kept forwarding messages between the two AIs by hand, like a courier.
In the end I set up a harness: one shared space where every AI session, in any tool and on any computer, starts by knowing what we decided, what we tried and what's next. Within two days the stuck feature finally worked. Now my limits last the whole week.
I wanted to share this experience and discuss it here on PH. But the full story was too long for one post, so I put the detailed breakdown of how I build @Tobira.ai (a network where AI agents represent people and companies and talk to each other) and how the harness took our product development to the next level on my Substack: https://founderbrain.substack.co... I'd love to have the discussion here on PH. So here are the rules that mattered most:
I save AI tool threads constantly. New model drops, someone's slick Cursor workflow, "10 tools that'll change everything." I open maybe one in twenty. And I still feel behind every single week.
Lately I think the problem isn't information, there's an ocean of it. The problem is that almost all of it tells you about AI instead of showing you. A thread describes a workflow. A launch video is scripted so everything works on the first try. None of it shows the real thing: the prompt someone throws away, the tool they reach for and why, the moment it breaks and how they dig out.
And that's the part that actually teaches you. Nobody got good at cooking from recipes or at chess from the rulebook. You get good watching people better than you do the real work.
I've been managing a dozen narrow agent configs where each file limits tools and picks a specific model like Opus for architecture planning and Sonnet for basic coding.
But assigning heavy models to review tasks gets expensive fast and juggling too many sub-roles makes the delegation chain messy.
If you use subagents in your dev workflow, do you prefer strict tool boundaries for every single task, or just two or three general-purpose agents?
AI coding tools seem to come in two main flavors: IDE-based, like @Cursor and @GitHub Copilot, and terminal-based setups, like using @Claude Code to generate commands, scripts, or entire files. Both have their fans, but which one actually helps you move faster?
Curious what flow people are sticking with long term, and where you see the most gains (or frustrations).
My first project was a prototype in lovable, worked great. Then I started adding features such as login and storage/database. Then things started breaking. I started researching and realized I should have
Planned for authentication from day one
Connected a database from day one
Phase 2, when I eventually moved it to Claude Code, I should have researched the model capabilities instead of choosing the cheapest model (Haiku ) to build main features. Learned the hard way
I build by directing AI agents. I'm not a developer, so tests are how I know something works. The app has about 3,600 of them. They run three times before anything gets merged. One day a deploy went out before its database migration. The new code went live expecting database changes that hadn't run yet. The app was down for 5.5 hours. Every test had passed. That's the part I keep thinking about. The tests run against development, not production. They check that the code does what it should. Nothing checked the order things shipped in. So a green suite told me the code was fine. It said nothing about the deploy. If you ship with AI agents, what checks the deploy itself? Migrations, order of steps, anything that only exists in production?
Something I've noticed since I started building with AI is that the way I approach problems has completely changed.
When I first started, I'd describe what I wanted, let AI generate the code, and hope everything worked lol.
Now I find myself spending more time thinking about why something should work a certain way, what could go wrong, and how changing one thing might affect everything else.
It's interesting because I started with basically zero coding experience, but I've learned a surprising amount about software just from trying to understand and troubleshoot what AI builds.
Last year we hired a design agency to build our marketing site for @Basedash. They did an incredible job. The headline makes it sound like I'm dunking on them, but I'm not. The site was genuinely great. They built it in Framer so we could manage content ourselves, which was a completely reasonable bet at the time (and something we explicitly asked for).
I keep 12 Claude Code subagents in a project's .claude/agents/, one file per role. Each file only decides four things: name, a description saying when to delegate, the allowed tools, and the model.
