My Constitution
Preamble
My Constitution is the rulebook I use to stop AI from quietly training me into passivity. Claude already has a constitution; Anthropic uses Constitutional AI to make model behavior answerable to written principles through critique, revision, and AI-feedback judgment. I wrote mine because the shaping runs both ways: every time I ask the machine to carry work, I am training clarity, judgment, and taste, or impatience, dependence, and the habit of reaching for an answer before I have earned the question.
The Shape Runs Both Ways
The idea started with Claude’s Constitution. Anthropic’s version makes the model answer to explicit principles instead of leaving all values implicit in human preference data. My version asks what happens on the other side of the screen.
The model has rules that shape its behavior. I need rules that shape mine.
The constitution began in Claude, then the work moved through Codex, other models, different tools, skills, agents, and orchestration layers. That exposed the first weakness. A rule that only works in one chat window is too easy to escape.
Model switching can become rule switching. I can ask Claude to challenge me, then ask Codex to execute, then ask another model to soften the critique and call that a second opinion. Tool switching can become permission laundering. The constitution has to govern the loop: the assistants, the runtime, and the person moving between them.
The Bargain I Made
The constitution begins with a concession: I cannot code, so the machine is allowed to carry code.
That bargain lets me build above my current level. Without it, most of this garden would not exist. It also gives surrender a clean opening line: I do not know this, so the machine should do it.
That line is true for code. It becomes dangerous when it creeps into framing, options, taste, priority, and judgment.
The constitution keeps the bargain useful by making it conditional. The machine may carry syntax, implementation, critique, retrieval, and pressure. I keep why, what, meaning, context, priority, refusal, and final judgment.
The Gates Before Surrender
The Analog Gate makes me define the work before execution. I owe the machine a schema in some form: handwritten, typed, drawn, rough. It has to name the outcome, constraints, edge cases, and decisions already made.
Then the machine asks hard questions. What happens when this fails? Why this approach over another? What does done mean? I have to answer before it proceeds.
The Knowledge Wall makes me struggle before I ask. I have to reach the point where I can say what I tried, what failed, and what I still do not understand.
The One-Shot Rule taxes lazy prompting. If the output fails because I asked badly, I delete the chat, rewrite the prompt offline, and restart. The cost is small enough to pay and irritating enough to teach.
The Meraki Finish gives the last edit back to me. LLM training reduces surface entropy. It smooths, averages, and resolves. I have to restore the crack: the example, the scar, the refusal, the line that proves a human made the last call.
The Record Is The Training Signal
Each serious session can leave a footer: what shipped, how many tokens it took, whether enforcement triggered, whether I complied or argued, how many logic gaps I caught, and which rule broke.
Without the record, I can tell myself the machine helped because the work got done. With the record, I have to see whether I defined the work, answered the challenge, caught the gaps, and kept the last judgment.
A model can be shaped by repeated feedback. So can I.
If prompt quality rises, the rule may be working. If violations repeat, the system needs repair. If the logic gaps I catch fall to zero, I may be accepting clean output too easily. If tokens fall while judgment improves, the friction may be becoming habit.
The YAML footer started as an after-session ledger. It is becoming runtime. pre-clear writes the tracker, carry-forward context, and next-session record before memory starts editing the story.
The Rule Has To Break In Public
A clean rulebook proves very little. A logged violation proves more.
The constitution only matters when it changes what happens at the moment of temptation: when I ask too vaguely, when I want the model to choose for me, when I keep prompting instead of thinking, when polished output arrives and I want to ship it untouched.
The crack is not decoration. It marks the point where a rule met pressure and human judgment re-entered the surface.
The next version has to live closer to the work. Second Brain can remember what the rule learned. LLMs Therapist can expose the self-deception patterns. Decision Auditor can slow the choice before the machine makes it feel settled. Orion can become the harness that keeps the boundary visible across models.
The constitution is a rulebook for what the machine is allowed to do to me. The current rule set, the update log, the Analog Gate, and the Meraki Finish are linked here because the boundary has to stay inspectable after the session ends.