Claude · Mechanism
How Claude Works
Claude is built on the same foundation as every other large language model, so the mechanism explained on our ChatGPT page applies here too. What is genuinely different is how it was trained to behave, and that difference shows up in how you should prompt it.
GPTPrompts.AI Editorial
Written from primary sources, with disagreements shown rather than hidden · Last updated August 2026
01
Constitutional AI, the actual difference
Standard alignment leans heavily on human raters comparing responses. That works, but it is expensive, inconsistent between raters, and hard to audit: the values end up implicit in a pile of preference data.
Anthropic's approach writes the principles down. The model is asked to critique its own response against that written constitution and revise it, and a large part of the feedback used for training is generated this way rather than by humans.
The practical consequence is that Claude's refusals and hedges tend to follow statable principles rather than a rater's mood, and Anthropic can publish what those principles are.
02
Why XML tags work so well
Claude is unusually responsive to explicit structural markers. Wrapping the parts of your prompt in tags, for example putting reference material inside a document tag and your task inside an instructions tag, reliably improves its handling of complex requests.
The reason is boundary clarity. When you paste a long contract and then ask a question, an untagged prompt leaves the model to infer where the material ends and the instruction begins. Tags remove that guess.
- Put pasted material inside its own tag.
- Keep instructions in a separate tag from context.
- Ask for the answer in a named tag when you need to parse it.
- Use example tags for few-shot patterns.
03
Long context, and what it does not fix
Claude handles long inputs well, which is why it is often the default choice for analysing lengthy documents in one pass rather than chopping them up.
A large context window is not perfect recall, though. Material in the middle of a very long input tends to get less attention than material at the start or end, an effect observed across long-context models generally. If a specific clause matters, quote it directly rather than trusting the model to find it.
Frequently asked questions
Is Claude better than ChatGPT?
They are close enough that the honest answer is task-dependent. Claude tends to be preferred for long-document analysis and sustained natural prose. ChatGPT has a broader product surface around it. The differences that matter most in practice are usually about which tools each is connected to, not raw model quality.
What is constitutional AI in plain terms?
Instead of teaching the model good behaviour purely by having humans rank answers, you give it a written set of principles and have it critique and rewrite its own responses against them. The values become explicit and auditable rather than implicit in preference data.
Can Claude browse the web?
Claude's capabilities depend on the surface you use and which tools are enabled, and these change. Do not assume live retrieval is happening. If an answer contains a specific claim about the current state of the world, check whether it actually retrieved anything or is generating from training data.
Does Claude hallucinate less than other models?
It is built on the same next-token mechanism, so it can and does fabricate. Alignment training changes how readily it hedges or declines, not whether fabrication is possible. Verify specifics from Claude exactly as you would from any other model.