Write with Claude. Review with Codex. New video: use a metaharness to perform a cross-repo cross-agent adversarial review. @trypolygraph #metaharness

I no longer think about many of my changes as bound to workspaces, PRs or repos. I think in prompts. I send a prompt and let the @trypolygraph figure out what needs to change, where, and in what order. Doesn't work for all changes, of course, but still super cool.

The government reapproved Fable because they hadn't heard Polygraph dropped and unleashed the full power. @trypolygraph trypolygraph.com/

I'm convinced that memory systems (also knowledge/brain) for code need a different approach than for softer domains like Slack, Notion, or email. Answers in Slack rarely have a true correct form. Correctness is filtered through the reader's interpretation. Code is different. It has correct answers. An "almost right" Slack answer still feels right. An "almost right" code answer is wrong, and demonstrably so. Think of an org chart note describing who owns what. It's never truly correct, just an approximation of what actually happens. That bar is fine for Slack. It's not fine for code. The upside is that code comes with far richer structured metadata (commits, dependency graphs). This metadata is deterministic and robust, which closes a lot of the gap. And unlike a Slack thread, code is still there. You can look at it, you can run it. The code itself doesn't drift, because it's what actually runs. Given that much higher bar for correctness, and the much richer toolset for hitting it, the solution for code memory will likely be different from the solutions for soft-knowledge memory.

A Genius With Amnesia My @aiDotEngineer keynote is out. youtube.com/watch?v=jVjt-2g8…