The Constitution

The Story

In April 2023, Samsung Electronics banned generative AI across multiple divisions after engineers had inadvertently leaked sensitive semiconductor data through an AI chatbot. The tool had been used for source-code optimisation and meeting-note summaries — reasonable use cases. The AI itself had no instructions about what information was confidential, no rules about what could not be shared externally, no constraints about data classification. It had no constitution.

The AI was helpful. It processed the semiconductor data, generated summaries, and assisted with code optimization -- exactly what it was asked to do. The problem: it did all of this without any awareness that the information it was handling was restricted intellectual property.

Samsung's response was to ban generative AI tools entirely for certain divisions. Not because the AI was broken. Because the AI had no constitution -- no identity, rules, or constraints that would cause it to recognize boundaries the way a human colleague instinctively would.

The difference between a ban and a boundary was a system prompt.

The Core Idea

Every AI product has a system prompt. It's loaded at the beginning of every context window, processed before the user's message, and applied to every response. Most teams treat it as a configuration setting -- a paragraph tossed in during development and rarely revisited.

That treatment radically underestimates what the system prompt is. It's not a configuration setting. It's the most important product artifact in any AI system. Every other CONTEXT layer varies per query. But the Constitution is constant. It shapes every response. It's the one piece of context that influences 100% of the model's outputs.

Instructions tell the model what to do: "Be helpful. Be accurate. Be concise." These restate defaults and change nothing. Constraints tell the model what to do when goals conflict: "When accuracy and speed conflict, sacrifice speed. When you don't have enough information, say so explicitly rather than guessing." These are the decisions that define the product.

Where This Hits in Production

Regulated industries and the auditable constitution. In healthcare, financial services, and legal, regulators increasingly ask: "How does your AI ensure it doesn't give harmful advice?" The constitution provides an auditable answer -- specific rules loaded into every inference call, in priority order.

Multi-tenant constitution management. In enterprise B2B, different customers may need different constitutions. A healthcare AI serving Hospital A (which allows AI-suggested medication alternatives) and Hospital B (which requires pharmacist approval) needs tenant-specific constitutions. A wrong constitution for the wrong tenant is a compliance failure.

Constitution drift. System prompts, like any code artifact, drift over time. New edge cases get patched with additional instructions. Rules accumulate without review. Six months after launch, the constitution has doubled in size and contains contradictory instructions.

In Practice: Anatomy of a Constitution

A production constitution is 2,800 tokens across 7 deliberate sections. Each one encodes a product decision, not a suggestion.

Remember This

1. Your system prompt IS your product definition. It's not a configuration setting -- it's the most important artifact in your AI product, shaping 100% of responses. Treat it with the same rigor as your product spec.

2. Instructions tell the model what to do. Constraints tell it what to do when goals conflict. Instructions that restate defaults add tokens without changing behavior. Write constraints.

3. The strongest constitutions have three properties: priority hierarchies (not wish lists), boundary definitions (not aspirations), and edge case protocols (not happy-path descriptions). Test them against your hardest cases.

Connecting the Dots

The constitution is the layer where product strategy meets AI behaviour. Every strategic decision -- who the product serves, what it values, what trade-offs it makes -- either shows up in the constitution or does not show up in the product.

The difference between Anthropic's Claude, OpenAI's ChatGPT, and Google's Gemini is not primarily a difference in model architecture. It is a difference in constitutional design. Claude tends toward refusal in ambiguous cases -- a constitutional decision prioritising safety and honesty over helpfulness in edge cases. ChatGPT tends toward helpfulness -- prioritising being useful over being cautious. Gemini tends toward groundedness -- referencing source material and available tools. None are wrong. They are product decisions, encoded in text the model reads 100% of the time.

A constitution that prioritises helpfulness over accuracy creates a product that users love and compliance teams fear. A constitution that prioritises accuracy over helpfulness creates a product that compliance teams approve and users find frustrating. The PM's job is to find the balance -- and encode it in the text the model reads before every response.

On the autonomy spectrum. At Levels 1-3, the constitution defines how the AI responds to each query. At Levels 5-7, the constitution defines the boundaries within which the agent operates without human review. The shift is profound: from respond this way to you may take any action within these constraints. The constitution becomes the governance framework for autonomous operation -- the rules the agent follows even when no human is watching.

Why this leads to the harness. A good constitution without a good harness is a document. A good constitution with a good harness is a product. The harness is what enforces the constitution at runtime: loading it consistently, managing its version, ensuring it sits at the highest-attention position in the context window, and detecting when behaviour deviates from constitutional rules. See What is a Harness.