Rules have three failure points.
The agent must read the rule, retain it, and choose to follow it. That is fragile even in one session and increasingly unreliable across many agents, long contexts, and repeated edits.
Field Note 04InfrastructureMarch 23, 2026
Operating principle
Rules to Infrastructure
Rules are suggestions. Infrastructure is physics.
When an agent must remember to obey a rule, the system has already placed the quality boundary in the wrong location.
Read the original on XThe evidence card
A concrete break is more useful than a general warning.
A growing CLAUDE.md file looked like discipline until it reached the point where compliance degraded. The durable move was not to write better instructions. It was to graduate repeatable rules into hooks, checks, and startup behavior that run whether the agent remembers them or not.
190lines before the second audit
84errors traced to one missing type entry
193perpetual animations found
0new anti-patterns after automatic scanning
The model
The translation layer.
The visual is an editorial compression of the source essay: enough structure to orient a reader, without pretending that the model replaces the underlying evidence.
InstructionInfrastructure
Run typechecks before finishing→Per-file typecheck after every edit
Read the campaign before starting→SessionStart loads durable state
Do not add repeat:Infinity→Post-edit anti-pattern scan
The argument
What changed when the break became visible.
The site version keeps the source's operational insight while making its sequence easier to inspect, share, and connect to the rest of the archive.
Move the boundary closer to the action.
A per-file typecheck after every edit catches the error at its source. A SessionStart hook loads campaign state before work begins. A regex scan warns on anti-patterns when they are introduced.
Keep the why. Automate the will.
Architecture, design philosophy, and project identity still belong in human-readable instructions. The moment a rule says always or never, ask whether the environment should enforce it instead.
The boundary
Useful systems name what they cannot prove.
Build physics around the decisions you cannot afford to forget.
Limit
Automation can enforce a known boundary; it cannot decide whether the boundary itself is wise, ethical, or still relevant.