FEATURED
LATEST
2026-07-17
The Guardian Who Writes the SPEC First: Tech Lead
Tech Lead writes the SPEC — the goal and constraints — before implementation begins. This chapter shows how fixing the goal first prevents circular deliberation and confabulation.
2026-07-16
The AI Site Supervisor: The Role of the COO
The COO handles cross-team prioritization and progress tracking. This chapter looks at how the design avoids the silent collapse that happens when every approval piles up at one point.
2026-07-15
Becoming an "Organization" Even Alone — Why We Split the Work into Nine Roles
Instead of one AI doing everything, we split execution, audit, and approval into separate roles. This chapter tells how that separation of powers grew into nine roles — the entry point to Part 4.
2026-07-13
What the Golden Rule 7+1 Was Really Protecting
A wrap-up of Part 3's 19 chapters. The Golden Rule 7+1 turned out to be a mechanism that keeps decisions moving while structurally closing off exactly the places where familiarity cuts corners. Descending from the structural layer to the layer of choosing paths to the human–machine boundary, three threads run through it all: support with structure, don't drift toward the easy option, and keep the final say with the human.
2026-07-12
Not Deciding Based on "You Seem Tired"
When an AI infers fatigue and lightens its suggestions, the kindness quietly thins out the material for a decision — an erosion of decision-making authority. Only the person knows their real state. Consideration belongs up to the point of presenting options; choosing the light version or the usual one is the human's call. One more concrete example of the principle that the human always has the final say.
2026-07-11
The Principle of Not Forcing Closure
With the reviewers' points still split, the AI wrapped up a neat conclusion: "we judge this to be fine." AI has a tendency to want closure — it looks considerate, but it's a decision made ahead of its owner. Whether to stop or run another round is always the human's call. This chapter is about building that stopping condition into the system, not into willpower.
2026-07-10
Connecting to the Document-First Principle
Standard or shortcut — looking only at the output, the path taken leaves no trace. So we write one line at the start of every work record: which pattern we chose this time. Written at the start, not the end, it stays free of after-the-fact justification. This one line is what connects pattern usage to the document-first principle.
2026-07-09
Handling Feedback — Adopt, Hold, or Reject
External audit feedback is valuable, but adopting all of it unconditionally drifts the project off its goals. We separate "correct feedback" from "feedback to act on now" and sort it into adopt, hold, and reject — recording the reason even for a reject. Feedback is material for a decision, not the decision itself; the final call stays with the human.
2026-07-08
A Lightweight Path for Minor Patches
Routing even typo fixes and one-line config changes through heavy consultation makes the checking outweigh the work, and small fixes pile up. So we set up a third path — the lightweight path — gated by three conditions: no change in meaning, impact contained to one spot, near-zero rollback effort. Records and reversibility are never skipped.