September Retrospective: What Ships Next

End of September. Time for the kind of retrospective that looks at the whole operation instead of any single project. Five months ago I started treating AI agents as first-class collaborators. Here's what I'd tell past-me about how it's going.

What's Clearly Working

Three things are now muscle memory: writing tests before implementation, keeping agents in folders with explicit contracts, and scoping work by verification time. These weren't intuitive in April. They're boring now. Boring is the goal.

The content pipeline ships. The client work ships. The skill soup thesis stopped being a thesis and became the default operating mode.

What's Clearly Breaking

Two patterns I'm still fighting:

1. Unverified work piles up when I go fast. The cap of one unverified change is correct. I break it under deadline pressure. I pay for it the next week.
2. Client onboarding still takes too long. Bringing a new client from intake to first shipped feature takes two weeks. It should take three days. The blocker is repeated context-setting that should be templated.

What October Needs to Prove

Three experiments for the next month:

• A templated client intake that compresses the two weeks to three days
• A harder cap enforcement on unverified work (maybe a git hook)
• One completely new tool added to the soup, end-to-end, on a real project

If those land, the operation gets another step function faster. If they don't, I need to rethink the bottleneck.

The One Thing I'd Tell Past-Me

Stop tweaking the stack. Pick tools that work, commit, then invest the freed attention in the work itself. April me spent too much time rearranging infrastructure. The compounding happens in the work, not the setup.