Deploy IntelligenceJuly 16, 2026
When AI Rules Should Become Code
I recently asked Claude Fable 5 to review my entire AI Constitution and help me understand why the system was starting to break around the edges. Its answer was interesting: the rules were not necessarily wrong, but too many decisions that should have been deterministic were still being left to language models.
My Constitution grew organically. Each time an AI did something unexpected, I added guidance intended to prevent it from happening again. This worked well when the system was smaller and the Constitution was primarily helping agents understand my preferences, projects, and ways of working.
But the system expanded across more agents, machines, accounts, providers, and services. Some of those rules stopped being guidance and became invariants.
Identity separation is one example. I maintain distinct personal and work identities on shared platforms. Mixing them can have serious consequences, so the Constitution tells every agent which identity belongs to each context.
That sounds like enforcement, but it isn’t.
If an LLM must determine the current context, remember the rule, inspect the active identity, and decide whether an operation is safe, the boundary remains probabilistic. Even a highly capable model can act on incomplete context, make a stale assumption, or simply miss a step.
A rule saying “never use the work identity in a personal repository” should therefore become tooling that:
- derives the required identity from the repository
- checks the authenticated account before any write
- refuses the operation when they do not match
The LLM no longer decides whether the rule applies. It cannot accidentally bypass it. The tooling makes the unsafe action unavailable.
The same principle applies elsewhere in my system:
- determine the actual execution host before attempting SSH
- prevent an agent from connecting back into the machine it is already running on
- record whether a path is synchronized by Obsidian Sync or Syncthing
- load credentials only from their canonical location and intended scope
- resolve current models from central policy rather than scattering versioned model names across prompts and jobs
These are not questions requiring intelligence. They are facts that software can verify.
The Constitution still matters. It should explain intent, establish boundaries, identify risk, and guide decisions that genuinely require judgment. But when a rule can be expressed as “before doing X, verify Y, and refuse if it is false,” it probably belongs in deterministic tooling.
This suggests a useful test for any AI operating system:
Are you asking the model to obey something that software could guarantee?
As these systems grow, improving the prompt or adding another constitutional rule will not always make them safer. Sometimes the Constitution has already done its job by identifying the invariant. The next step is to turn that invariant into code.