Teaching Machines to Teach Machines
September 21
“We do these things not because they are easy, but because we thought they would be easy!”
This Kennedy quote parody which showed up in both the ODI and 18F Slacks at variouss times, came to mind around hour ten of this long Sunday session, as I wrapped up what might be the most meta work I’ve ever done: spending hours creating briefing documents that teach AI agents how to onboard other AI agents to build an AI assistant.
If that sentence made your head spin a little, welcome to the recursive reality of modern development.
The problem nobody warned me about
When I started building Piper Morgan with AI collaboration, I focused on the obvious challenges: Can Claude write good code? Can Cursor handle complex refactors? How do I coordinate multiple agents effectively?
What I didn’t anticipate was the briefing problem.
Every new chat with an AI agent starts at zero context. They don’t remember yesterday’s architectural decisions, last week’s failed approaches, or the hard-won lessons from three months of systematic development. Each session begins with the same routine: “Here’s what we’re building, here’s how we work, here’s where we are…”
By September, I was spending more time re-explaining context than actually building. Worse, different agents were getting slightly different context each time, leading to inconsistent approaches and confused handoffs.
The 75% pattern kept recurring — agents would start working on solutions that looked right but didn’t fit our established patterns, or they’d suggest “improvements” we’d already tried and abandoned.
Enter the “briefing” concept
Last weekend, I faced a choice: start CORE-GREAT-1 (our first major refactor) with the usual context scramble, or take time to build systematic onboarding infrastructure.
Experience told me that momentum favors the former. Product management wisdom suggested the latter.
I chose boring.
Twelve hours later, I had six briefing documents that form a four-level onboarding system:
Level 0: Methodology — How we work (Inchworm Protocol, Excellence Flywheel)
Level 1: Project Context — What we’re building (Piper Morgan’s vision and technical foundation)
Level 2: Current State — Where we are right now (current epic, recent progress, known blockers)
Level 3: Role-Specific — Your specific responsibilities (Chief Architect vs Lead Developer vs Programmer)
Now every agent gets role-appropriate context without me re-explaining the same architectural decisions or methodology principles.
Recursive intelligence recursion
But here’s where it gets properly meta: I used AI agents to design the briefing system for AI agents.
Claude helped architect the four-level framework. Another Claude instance wrote the methodology documentation. A third reviewed the role-specific guidelines for consistency.
We were literally teaching machines to teach machines how to collaborate systematically.
This stuff is my happy place, to be honest. Hey, look, there’s a reason why my band is called Layers of Meta.
Making the mesh tighter
Will this prevent all context loss and coordination confusion?
Ha! Experience tells me that chaos will still find a way in, but we keep trying to make the mesh tighter.
What the briefing infrastructure does provide:
- Consistent starting points for every agent session
- Role-appropriate depth so agents get context they need without noise they don’t
- Systematic handoffs between different AI collaborators
- Embedded lessons learned so we stop repeating solved problems
More importantly, it creates a foundation for scaling AI collaboration. Instead of my time being the bottleneck for context transfer, agents can onboard other agents using systematically designed materials.
The infrastructure invisible
From the outside, yesterday’s work looks like the least productive session imaginable. No features shipped. No bugs fixed. No user-facing improvements.
Just 12 hours of writing documentation that teaches machines how to read documentation.
But this is exactly the kind of infrastructure work that distinguishes sustainable development from heroic sprints. The briefing system becomes invisible when it’s working — future sessions will start faster, stay focused longer, and hand off cleaner.
The true test comes Monday, when CORE-GREAT-1 begins with every role properly briefed and aligned.
What this means for AI-augmented development
If you’re building anything substantial with AI collaboration, you’ll eventually face the same briefing challenge. The tools are powerful, but they’re amnesiac. Success requires as much systematic thinking about knowledge transfer as about the product itself.
Some emerging patterns I’m seeing:
Documentation for machines is different than documentation for humans. Agents need explicit context about what NOT to do, embedded reminders about methodology, and role-specific filtering of information.
Onboarding systems need onboarding systems. The meta-level quickly becomes recursive — you need AI help to design effective AI collaboration, which requires briefing the AI that designs the briefings.
Infrastructure timing is everything. Too early and you’re optimizing for problems you don’t understand yet. Too late and you’re drowning in context debt.
The boring work enables the interesting work. Systematic foundations create space for creative problem-solving instead of constant re-explanation.
The “Ask not” moment
Around hour 10, as I finished the sixth briefing document and updated the agent prompting templates, that gag hit me: “We do these things not because they are easy, but because we thought they would be easy!” It’s funny because it’s true.
Building an intelligent PM assistant seemed straightforward enough. Building the systematic infrastructure to coordinate multiple AI agents to build an intelligent PM assistant while maintaining consistency and preventing knowledge loss?
That’s a different problem entirely.
But here’s the thing — it’s exactly the kind of problem that becomes invisible when solved correctly. Monday’s session will start with proper briefings, clear roles, and systematic context. The work will flow better because the infrastructure is in place.
And maybe, just maybe, we’ll make the mesh a little tighter against the chaos that’s always trying to find its way in.