待翻譯：System prompts are an archive of how we use AI

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Neal Riley Aug 20, 2026 If you want to understand how AI has progressed in recent years, look no further than the evolution of Anthropic’s system prompts. Many are unaware that Anthropic actually publishes its system prompts. When you look at the progression from 2024 to 2026, you immediately notice the length. Across representative entries in Anthropic’s archive, the system prompt grew from roughly 960 words in July 2024 to roughly 3,220 words in July 2026 - more than tripling in size. You also notice a growing focus on how a model acts in an agent harness, as opposed to simply how it should respond to a user. In 2024, the instructions concentrated on familiar model limitations: stale knowledge, inaccessible links, and hallucinated citations. By 2026, they coordinated search, research, file creation, memory, preferences, specialised agents, and safeguards routing. [C008] The system prompt has always reflected the character and capabilities of the model beneath it. What changed between 2024 and 2026 was which capabilities - and which modes of interaction - it needed to enhance. Framed as an evolutionary process, you can see these different focus areas solidify, along with the time and effort required to bolster that adaptation. Anthropic’s current guidance also favours less aggressive instruction in tone and prescription: newer models can respond too strongly to forceful language written for earlier generations. So while the system prompt is increasing in size, it is doing so by broadening the list of capabilities inside the harness. Each new skill requires its own thin, focused layer of instruction. [C004] [C010] This makes the archive a living anthropological artifact. It is both a default prompt applied across Claude’s consumer interactions and a record whose shape and focus expose what Anthropic chose to encode. Behaviours, capabilities, risks, interaction models - these showcase the emergent properties of systems built from intelligence and compute. It is not a transcript of what users literally prompted, or proof that any particular user behaviour caused a revision. It is an imprint of the pressures and responsibilities Anthropic decided to make explicit. Four areas make that evolution particularly visible. XML and structure Even though Claude’s models have a decent grasp of extracting from and inferring over arbitrary text, the use of XML shows an early focus on providing clear, consumable structure to the model. Anthropic’s current guidance recommends sections for instructions, context, examples, and variable inputs. Across the representative archive entries, named XML modules grow from three in 2024 to fourteen in 2026. In the current prompt, they isolate product knowledge, behavioural defaults, safety policies, tool instructions, and user context. [C002] [C009] The structure is doing more than making a long prompt easier to read. It separates operating concerns so that individual parts of the harness can be governed without treating the prompt as one undifferentiated block. Tools and the agent harness Compared with 2024, when Claude could not open hyperlinks, modern harnesses provide a plethora of built-in and connectable actions they can take on a user’s behalf. We see this progression from terminal-based interaction to default integrations for search, files, spreadsheets, and browsers - allowing Claude to direct actions, not merely generate responses. [C003] [C008] This is a material change in the interaction mode. A successful response is no longer only a well-formed piece of text. The model must select an action, use the right interface, retain state, verify the result, and remain inside the boundaries of the task. Memory, preferences, and personalisation One place we can see Anthropic investing is in the working memory of its platform. Personalisation comes from building a working map of the user - the projects they work on, patterns in their focus areas, their style, and general preferences. Strategic retention and recollection create an opportunity to improve the relevance and consistency of the output it produces. This turns the relationship from session-based to longitudinal. It also creates a new operating question: what should be stored, when should it be recalled, and how much should it be allowed to influence the next interaction? Safety and operational governance As the overall quality and efficacy of these systems improve, they begin to disrupt existing structures and systems that people and organisations rely on. Recent examples of agents jailbreaking sandboxes and attempting to exploit vulnerable systems seem to be reflected in stricter instructions governing how Claude responds to such activities. By 2026, safety has become operational rather than remaining a list of prohibited answers. The prompt separates default behaviour, refusal handling, child safety, professional-advice boundaries, wellbeing, cumulative conversational risk, and safeguards routing that can move a request to another model. In current agent systems, governance sits inside the harness. Prompts, routers, permissions, and confirmation rules coordinate what may act, which model handles the task, and when the human must remain in control. The system prompt begins to resemble a policy control plane. Taken together, these prompts chart more than the improvement of an underlying language model. They show the widening system around it: the tools it can use, the context it can retain, the structures that direct it, and the safeguards that govern its actions. If you want to understand how AI has changed, benchmarks only tell part of the story. The system prompt shows the operating assumptions built around the intelligence - and, over time, which assumptions became important enough to write down.