Shaping the flow of ideas at the frontier of cognitive systems.
—
Between 2023 and mid-2025, Claude’s enterprise market share climbed from twelve percent to thirty-two percent. Over the same period, OpenAI’s share dropped from fifty percent to twenty-five. The shift is even more pronounced in coding workloads: Claude now commands forty-two percent of enterprise use—more than double OpenAI’s share. When organizations spend real money on production systems, they are choosing one approach over another. The question is: what are they choosing?
In January 2026, Anthropic published its eighty-page Model Specification—what the company calls Claude’s Constitution. Past the headlines about consciousness speculation, the document reveals a design bet: teaching an AI system why to behave produces better results than telling it what to do. The enterprise numbers suggest the bet is paying off. The Constitution describes this as training Claude to behave like “a thoughtful senior Anthropic employee”—someone who doesn’t just follow instructions, but understands purposes and can reason through situations the rulebook never anticipated.
The practical difference shows up in how Claude handles gaps. When instructions don’t cover a scenario, most AI systems either refuse to act or default to rigid fallback behaviors. Claude fills gaps with judgment. A customer service agent built with Claude, instructed to help users with a software product, will also answer general coding questions even if the system prompt never mentioned them. It infers the spirit of what the operator wanted rather than halting at the edge of explicit rules. This gap-filling behavior is not a bug or a workaround—it is the core of how the system is designed to work.
—
This changes what belongs in a system prompt—the instructions that developers give to Claude when building applications. Most developers write system prompts as configuration files: lists of rules, constraints, and handlers for edge cases. The Constitution suggests a different approach: write them as onboarding documents.
Consider two ways to handle the same task. A rule-based prompt might say: “Never discuss competitor products. If asked about competitors, say you can only discuss our products.” A reasoning-based prompt might say: “You’re representing Acme Corp in customer conversations. We want customers to have a helpful experience focused on whether our product solves their problem. Discussing competitors in detail would shift the conversation away from understanding the customer’s needs. If competitors come up, acknowledge the question and redirect to understanding what the customer is trying to accomplish.”
The reasoning-based version is longer, but it handles edge cases better. What happens when a customer says, “I’m evaluating you against Competitor X—can you help me understand the differences?” The rule forces awkward deflection. The reasoning allows helpful engagement while staying aligned with the actual goal. Claude follows rules competently. It follows reasoning excellently. The difference compounds in production at scale.
The onboarding document approach starts with context, not commands. Explain the situation. Describe what success looks like. Articulate which constraints matter and why. Add specific rules only where Claude’s default judgment diverges from what you need. The model is trained to act like a professional who has been briefed on a project. Write the brief accordingly.
—
The most ambitious claim in the Constitution appears in its treatment of autonomous agents—AI systems that can act on a user’s behalf without constant supervision. Most agents today operate like bureaucrats: they follow workflows, execute predetermined steps, and halt when they encounter situations their designers didn’t anticipate. Their value is capped by what the builder could specify in advance.
The Constitution describes something different: an entity with phronesis, Aristotle’s term for practical wisdom. Not just knowledge of rules, but the disposition to reason well when rules run out—the capacity to discern the right action in particular circumstances.
Imagine an agent managing a calendar, given the instruction to “prefer mornings for focus time.” A rigid agent, confronted with a VIP customer who emails at four in the afternoon requesting an urgent call tomorrow morning, halts. It recognizes a conflict but has no framework for resolving it. An agent with practical wisdom understands why mornings are preferred—for deep work and energy—and why the VIP matters—for revenue and relationship. It weighs these purposes against the specific context, makes a reasonable call, and explains its reasoning.
This changes agent architecture in three ways. First, designs can simplify: describe goals and constraints, then let the agent navigate toward them, rather than building rigid state machines with explicit decision trees for every possible branch. Second, evaluation becomes harder and more important—you cannot unit-test judgment; you need scenario-based frameworks that probe how an agent handles ambiguity and conflicting priorities. Third, agent prompts should explain purposes, not just specify behaviors. An agent that understands why you care about customer response time will handle novel situations better than one that only knows the service-level agreement number.
—
The Constitution is Anthropic’s bet on building AI systems we might trust to act on our behalf in genuinely complex situations. The enterprise market validates the approach at production scale. But the deeper shift is architectural. Building systems with practical wisdom requires different design principles than building rule-following automata: context over configuration, reasoning over enumeration, purpose over procedure. The models are learning judgment. The question is whether we are building the infrastructure to use it.
—
Source: “My breakdown of Claude’s 80-Page Constitution + 3 prompts to use it properly” — Nate’s Newsletter (natesnewsletter.substack.com), February 6, 2026.
Original to this post: The VIP customer calendar scenario (illustrating phronesis) is constructed to clarify the concept; the Constitution discusses phronesis but does not provide this specific example.
Text summarized and optimized using Anthropic’s models and reviewed by a human.