📝 SUMMARY This week's call centered on the evolution from single AI assistants to "chief of staff" multi-agent architectures, with multiple members converging on similar two-tier systems where a coordinator agent delegates to an implementer via shared repositories. Patrick Chouinard detailed his enterprise rollout of 2000+ Claude licenses and the governance patterns required, while others shared strategies for managing model verbosity, measuring real ROI versus vanity metrics, and overcoming organizational resistance to AI coding. Additional discussions covered database migrations for real-time apps, personal assistant tooling, and enterprise adoption tactics. 💡 KEY INSIGHTS The "assistant" model collapses at scale. Patrick described how coordinating many Claude assistants eventually consumes more time than it saves, leading his team to create a two-tier architecture: Claude Cowork acts as "chief of staff" coordinating with Claude Code as "implementer" through a shared local Git ledger, reducing humans to pure decision-makers. Two agents talking directly is the real unlock. Patrick and Paul Miller independently arrived at architectures where coordinator and executor agents communicate and write status to each other without human message relay, freeing the human to act only as a decision-maker. Surface decisions and wins, not just noise. Ty Wells noted that showing only problems creates a demoralizing interface. Effective chief-of-staff agents should highlight progress and successes alongside escalations, not just raw agent output. Set explicit time budgets. Paul Miller emphasized that agent tasks without deadlines can silently balloon from one hour to four, cascading delays. Decisive time constraints prevent runaway context windows. Opus 5 requires verbosity management. Multiple attendees flagged Opus 5 as unusually chatty. Mitigations include Matt Pocock's "wait, what?" skill to detect confusing output and Patrick's discovery that sarcastic, personality-driven prompts naturally produce shorter, clearer responses.

As promised during this week’s coaching call, and at the request of several participants, I’ve published the meta prompt we discussed as a GitHub Gist. The goal is to turn a frontier reasoning model like Claude Fable into a strategic analyst instead of just a coding assistant. Rather than focusing on a single project, it analyzes an entire development corpus, mines previous AI sessions and memory, reconstructs intent across projects, identifies hidden opportunities, and identifies the handful of problems that are actually worth spending frontier-model tokens on. If you’re juggling multiple AI projects, agent frameworks, research initiatives, or a growing codebase, I think you’ll find it useful, or at the very least it’ll give you ideas for building your own version. All values in <%YOUR VALUE HERE%> need to be replaced by your real information. !!! WARNING I altered the prompt to target 'high' or 'xhigh' Effort in order to try to contain the token usage, but when I ran it, I ran it on UltraCode. ONLY do that if you are ready to spend an entire 5-hour reset limit on a single prompt run. On UltraCode this is insanely token hungry WARNING!!! GitHub Gist: https://gist.github.com/hopchouinard/60d4d6e0d477e22d344ef75489fb2149 If you improve it or adapt it to another model, I’d genuinely be interested in seeing where you take it. The prompt is only half the idea. The methodology behind it is what I hope proves useful over time.

Live tomorrow 2 PM ET: watch builders demo what they shipped with Claude Code Tomorrow (Wednesday) at 2 PM ET we run our weekly Show & Tell. Builders get about 2 minutes to demo something they actually shipped with Claude Code or Codex, then about 5 minutes of honest feedback from the room. Real builds, real feedback, no slides. Free to watch, no signup: https://shipsafe.franklabs.io/watch That page can also email you the live link an hour before we start. Nothing else, unsubscribe any time. If you want to see how other people run Claude Code and Codex day to day, it is a good hour. Come lurk.

Deep in a session, right when I understood the work best, I wrote down what to build next as a short structured capsule. Then I cleared the context. A fresh session with none of that understanding read the capsule cold and built it to spec. It did not know it was finishing its own work. Handoff notes fail two ways: they ROT (you forget) or they CANNOT TRAVEL (a fresh session can't read your mind). And terse notes silently drop the two things that matter most: WHERE the work goes and what DONE looks like. I measured it. Labeled-field capsules scored 10/10 on intent fidelity vs terse prose at 9.67, and prose's weakest spots were exactly WHERE and ACCEPTANCE. The honest part: the model still ran ~31% of malformed input instead of refusing it, so that check is deterministic code, not a model call. It understands the capsule. It does not get to decide the capsule is safe. Free CLI and Claude Code plugin, MIT. Try authoring one capsule next time you are deep in something, then clear and let a fresh session build it. github.com/gtsbahamas/intent-capsule

OpenClaw, NanoClaw, ClaudeClaw. An assistant in your chat that reads your inbox, remembers everything, and gets things done. Lean in. The people getting the most out of them aren't the most careful. They just wear a seatbelt, so they can drive faster: - Read-only to start. Let it earn more. - It asks before it sends, deletes, or pays. - Its own space, not your main account. That's most of the protection, at zero cost to what it can do. Full playbook, plain English, no login. Plus a prompt your agent runs to audit itself: https://shipsafe.franklabs.io/agent-safety Which one are you running, and what have you got it doing?