Theo spent $1,500 in tokens trimming T3 Code's tools
Theo built an 80-case eval from his own rage moments with T3 Code's tools, and Opus failed over half of them. The fix cost $1,500 in tokens and cut tool context by 79%.
The big disagreement
Theo says Grok 4.7 should not be used, and day-one fans should not be trusted
Theo posted that Grok 4.7 probably should not be used at this point, with a link rather than a written-out test. Pressed in a reply, he went further: anyone who called it "best model ever" at launch should not be trusted. He does not say what he ran.
Anyone who thought Grok 4.7 was "best model ever" at launch should not be trusted
Receipts
Theo's tool-description eval: 4x better tool use, 79% less context
Theo says agents now use T3 Code's tools 4x better while using 79% less context, after a week spent building an eval out of his day to day tool usage and his failure cases.
Amazed at how many problems you can kinda just turn into evals and then do shit like this.
How they actually work
Pocock splits one-way-door PRs into a revertible half and a cutover
Matt Pocock's prompt of the day asks the agent to split a "one-way door" PR into two: one that implements everything but can be easily reverted, and a stacked followup that actually goes through the door. He calls it expand/contract, with feature flags as needed. The point is testing large chunks in production before the irreversible step.