The Honest Machine · Lesson 6
What you'll be able to do: place any piece of the build system — a file, a rule, a skill, a machine — on the correct layer, and trace how a lesson learned in one project reaches every future project through the flywheel.
| Layer | Name | What it is | Example |
|---|---|---|---|
| 1 | Habitat | Any machine, treated as replaceable | home PC, a VM, a second laptop |
| 2 | Project | One repo per project, born with its chassis | a web tool, a coach app, a game project |
| 3 | Mesh | How work moves: roles, laps, branch→PR→CI→human gate | one agent orchestrates, another executes, I merge |
| 4 | Harness | Cross-project doctrine every project inherits | skills, START_HERE, Knowledge_Base, diary, scheduled sweeps |
And the flywheel that connects them: lessons flow up (a hard-won lesson in any project gets packaged into the harness as a skill or doctrine), and doctrine flows down (every fresh session inherits the harness, zero re-explaining). The whole point: memory lives in the builds, not in any one chat, machine, or vendor.
A school district. Habitat = the buildings — a school is not its building; if one floods, classes move. Project = each course, with its own binder and gradebook (Lesson 1's chassis). Mesh = the schedule and staffing rules — who teaches, who substitutes, who signs off on grade changes. Harness = the district's teaching philosophy and teacher-training program: no single course owns it, every course inherits it, and when one teacher discovers a technique that works, the training program absorbs it and every future classroom gets it.
Where the analogy ends: a district's philosophy spreads at human speed — workshops, years. The harness spreads at session speed: a lesson packaged as a skill today is enforced behavior in tomorrow's first session. And unlike a district, layers here are swappable in both directions: not just new buildings, but in principle a new AI (the model-agnostic claim — which the stack map honestly labels a claim, not a proof: only Claude has run the full harness so far).
Sort these onto layers 1–4 (answers at the bottom of the lesson; no peeking until committed):
RESUME.md in the repo of a tool I builtSTART_HERE.mdThen walk one real flywheel loop: find a lesson in this tier that started life inside one repo (Lesson 3's law, in a tool I built) and name where it lives now (this lesson set — harness).
▲ lessons up doctrine down ▼
L4 HARNESS ── skills, START_HERE, KB, diary ── inherited by every session
L3 MESH ── roles, laps, PR→CI→human gate ── how any work moves
L2 PROJECT ── repo + chassis, born together ── where work lives
L1 HABITAT ── disposable machines, VMs ── where work runs
Two design insights. First, each layer protects the one above from the one below: projects don't care which machine (L1 is cattle); the mesh doesn't care which project (same loop everywhere); the harness doesn't care which agent (doctrine is text). Second, the harness is the flywheel's hub, which makes it the highest-value target — the adversarial review caught exactly that: the most-leveraged layer had the least protection (no git, no backup). The fix was a private repo for the shared doctrine — the harness now wears the same chassis it prescribes. A system that exempts itself from its own laws isn't a system yet.
Builder answers: 1→L2, 2→L4, 3→L3, 4→L1, 5→L4, 6→L2 (enforced by L2's CI; the pattern it teaches lives in L4), 7→L4.
Lesson complete — you can now see the whole machine at once. Next: Lesson 7 — what happens when a layer claims a safety net it doesn't have.