Claude Is a Product With a Constitution

Thesis

Claude is a product with a constitution. The principle list is real, published, and used in training. The object you pay for is still claude.ai and an API: Projects, Artifacts, a system prompt, a usage bar. Treat the constitution as a control surface on that product. Do not treat it as a court.

Context

I use Claude from a desk in Auckland for drafting and for a second pass on mystic-bytes copy. I use the API when I want a model ID instead of a chat habit. I have never been on Anthropic’s staff. There is a separate essay about the firm’s bet: Constitutional AI as a paper, a public principle list, an RSP. This one is the product shape those papers funded. What you click. What you get billed for. How a refusal arrives in the UI.

“Helpful, honest, harmless” is a slogan that fits on a plan page. The plan page is the document I actually renew.

Mechanism

claude.ai is a workspace, not a personality. Projects, artifacts, a file list, a chat that can draw a page in a side panel: this is product design for people who want an assistant that keeps context in a folder.¹ I put a style note in a Project and watch the next reply lean toward it. That is retrieval and instructions, the same machinery every vendor is shipping. The constitution is upstream of the tone. The folder is why I came back. Confusing those two is how you end up defending a brand when a Project just lost your thread at the usage cap.

The API is the contract surface. Messages, a system prompt you supply, tools, a model ID, rate limits. I prefer it when the work might ship, because a chat UI will change and an HTTP call is greppable. Anthropic publishes model cards and, unusually, snapshots of the system prompts that steer the consumer assistant.² Read those prompts as product source. They are long. They include the house manners, the refusal style, the bits that make Claude sound like Claude. A constitution that only lived in a 2022 paper would be a research story. A system prompt you can read is a product claim you can diff when they update it.

Constitutional behavior is UX. The papers described a model that explains a refusal instead of stonewalling. In the product that shows up as a particular kind of no: polite, essayistic, sometimes more helpful than you wanted, sometimes less useful than a competitor that just did the thing. I have hit both. A constitution that over-weights caution wastes a turn. One that over-weights helpfulness will polish a bad idea until it sounds shippable. The product team picks a point on that curve every release. You feel it as vibe. It is a setting they do not expose as a slider.

Plans are the other constitution. Free, Pro, Team, API prepaid: different caps, different model pickers, different whether Artifacts and longer context feel like a feature or a tease. When a reply stops mid-thought because the bar ran out, the principle list did not fail. The SKU did what SKUs do. I keep a small budget and I keep the important drafts in git. Claude is allowed to vanish at a cap. mystic-bytes is not.

Safety copy is part of the purchase. You are buying an assistant that will sometimes lecture. That lecture is cheaper than a lawsuit and more marketable than a raw model. I do not resent it. I also do not confuse it with aligned judgment about whether a Dark Heart Labs sentence sounds like me. The constitution does not know my voice. The Project might, if I pasted enough. I still read the output.

Tradeoffs

Refusal style vs getting the work done. An explained no is better manners than a silent filter. A no on a lawful draft I needed is a wasted seat. If the product refuses work you can do locally, do it locally. Do not argue with a control surface as if it were a colleague.

Chat vs API. Chat is right for artifacts and a Project folder. API is right for anything with a deadline and a model pin. I use both. I do not let the chat become the system of record.

When a different SKU is correct. If you want a sloppier, faster yes, another vendor will sell it. If you want the principle list and a published system prompt, Claude is a coherent product. Pay for that. Quote the prompt. Do not baptize the chat.

Close

Hold both: the constitution is why this product feels like this product, and the product is still a usage bar with a folder. I will keep Claude where it is the right SKU. I will keep the last edit. A principle list you can read is a gift. A gift is not a mind, and it is not a court. It is part of the software.

— JV · Dark Heart Labs.

References

Anthropic, Claude product documentation (claude.ai Projects, Artifacts, plans). The workspace and the seat, as distinct from the 2022 paper. ↩
Anthropic, published Claude system prompts and Claude API documentation (Messages, models, rate limits). Product source you can read, plus the HTTP contract. ↩