This text was initially revealed as two posts on Drew Breunig’s weblog. He’s been sort sufficient to share them right here.
Again in Might, Ásgeir Thor Johnson satisfied Claude to surrender its system immediate. The immediate is an efficient reminder that chatbots are extra than simply their mannequin. They’re instruments and directions that accrue and are honed by person suggestions and design.
For individuals who don’t know, a system immediate is a (typically) fixed immediate that tells an LLM the way it ought to reply to a person’s immediate. A system immediate is sort of just like the “settings” or “preferences” for an LLM. It’d describe the tone it ought to reply with, outline instruments it might probably use to reply the person’s immediate, set contextual info not within the coaching information, and extra.
Claude’s system immediate is lengthy. It’s 16,739 phrases, or 110 KB. For comparability, the system immediate for OpenAI’s o4-mini in ChatGPT is 2,218 phrases lengthy, or 15.1 KB—~13% the size of Claude’s.
Right here’s what’s in Claude’s immediate:
Let’s break down every part.
Device definitions
The most important part, the Device Definitions, is populated by info from MCP servers. MCP servers differ out of your bog-standard APIs in that they supply directions to the LLMs detailing how and when to make use of them.
On this immediate, there are 14 totally different instruments detailed by MCPs. Right here’s an instance of 1:
This instance is straightforward and has a really brief “description” subject. The Google Drive search device, for instance, has an outline over 1,700 phrases lengthy. It could actually get advanced.
Different device use directions
Exterior the Device Definitions part, there are a lot extra device use directions—the Quotation Directions, Artifacts Directions, Search Directions, and Google Integration Watchouts all element how these instruments needs to be used throughout the context of a chatbot interplay. For instance, there are repeated notes reminding Claude to not use the search device for subjects it already is aware of about. (You get the sense that is/was a troublesome conduct to remove!)
In reality, all through this immediate are bits and items that really feel like hotfixes. The Google Integration Watchouts part (which I’m labeling; it lacks any XML delineation or group) is simply 5 strains dropped in with none construction. Every line appears designed to dial in preferrred conduct. For instance:
If you’re utilizing any gmail instruments and the person has instructed you to search out messages for a selected individual, do NOT assume that individual’s electronic mail. Since some staff and colleagues share first names, DO NOT assume the one that the person is referring to shares the identical electronic mail as somebody who shares that colleague’s first identify that you will have seen by the way (e.g. by a earlier electronic mail or calendar search). As a substitute, you may search the person’s electronic mail with the primary identify after which ask the person to substantiate if any of the returned emails are the right emails for his or her colleagues.
All in, practically 80% of this immediate pertains to instruments—easy methods to use them and when to make use of them.My instant query, after realizing this, was, “Why are there so many device directions exterior the MCP-provided part?” (The grey bins above.) Poring over this, I’m of the thoughts that it’s simply separation of issues. The MCP particulars include info related to any program utilizing a given device, whereas the non-MCP bits of the immediate present particulars particular solely to the chatbot utility, permitting the MCPs for use by a number of various functions with out modification. It’s customary program design, utilized to prompting.
Claude conduct
On the finish of the immediate, we enter what I name the Claude Conduct part. This half particulars how Claude ought to behave, reply to person requests, and prescribes what it ought to and shouldn’t do. Studying it straight by evokes Radiohead’s “Fitter Happier.” It’s what most individuals consider after they consider system prompts.
However hotfixes are obvious right here as nicely. There are lots of strains clearly written to foil frequent LLM “gotchas,” like:
- “If Claude is requested to rely phrases, letters, and characters, it thinks step-by-step earlier than answering the individual. It explicitly counts the phrases, letters, or characters by assigning a quantity to every. It solely solutions the individual as soon as it has carried out this express counting step.” This can be a hedge in opposition to the “What number of R’s are within the phrase ‘Raspberry’?” query and comparable stumpers.
- “If Claude is proven a traditional puzzle, earlier than continuing, it quotes each constraint or premise from the individual’s message phrase for phrase earlier than inside citation marks to substantiate it’s not coping with a brand new variant.” A standard solution to foil LLMs is to barely change a standard logic puzzle. The LLM will match it contextually to the extra frequent variant and miss the edit.
- “Donald Trump is the present president of the US and was inaugurated on January 20, 2025.” Based on this immediate, Claude’s data cutoff is October 2024, so it wouldn’t know this truth.
However my favourite word is that this one: “If requested to jot down poetry, Claude avoids utilizing hackneyed imagery or metaphors or predictable rhyming schemes.”
Studying by the immediate, I’m wondering how that is managed at Anthropic. An irony of prompts is that whereas they’re readable by anybody, they’re troublesome to scan and normally lack construction. Anthropic makes heavy use of XML-style tags to mitigate this nature (one has to surprise if these are extra helpful for the people enhancing the immediate or the LLM…) and their MCP invention and adoption is clearly an asset.
However what software program are they utilizing to model this? Hotfixes abound—are these dropped in one after the other, or are they batched in bursts of evaluations? Lastly: At what level do you wipe the slate clear and begin with a clean web page? Do you ever?
A immediate like this can be a good reminder that chatbots are way more than only a mannequin, and we’re studying easy methods to handle prompts as we go. Fortunately, Ásgeir Thor Johnson continues to gather these prompts in a GitHub repository, permitting us all to simply observe alongside. And following adjustments made to those prompts—which you are able to do by reviewing the historical past of Johnson’s repo—renders their improvement extra clear.
Claude’s system immediate adjustments reveal Anthropic’s priorities
Claude 4’s system immediate is very just like the 3.7 immediate we analyze above. They’re practically equivalent, however the adjustments scattered all through reveal a lot about how Anthropic is utilizing system prompts to outline their functions (particularly their UX) and the way the prompts match into their improvement cycle.
Let’s step by the notable adjustments.
Outdated hotfixes are gone; new hotfixes start
We theorize above that many random directions concentrating on frequent LLM “gotchas” have been hotfixes: brief directions to deal with undesired conduct previous to a extra sturdy repair. Claude 4.0’s system immediate validates this speculation—all the three.7 hotfixes have been eliminated. Nevertheless, if we immediate Claude with one of many “gotchas” (“What number of R’s are in Strawberry?” for instance) it doesn’t fall for the trick. The three.7 hotfix behaviors are virtually actually being addressed throughout 4.0’s posttraining by reinforcement studying.
When the brand new mannequin is educated to keep away from “hackneyed imagery” in its poetry and assume step-by-step when counting phrases or letters, there’s no want for a system immediate repair.
As soon as 4.0’s coaching is completed, new points will emerge that have to be addressed by the system immediate. For instance, right here’s a brand-new instruction in Sonnet 4.0’s system immediate:
Claude by no means begins its response by saying a query or thought or statement was good, nice, fascinating, profound, glorious, or some other optimistic adjective. It skips the flattery and responds immediately.
This hotfix is clearly impressed by OpenAI’s “sychophant-y” GPT-4o flub. This misstep occurred too late for the Anthropic crew to conduct new coaching concentrating on this conduct. So into the system immediate it goes!
Search is now inspired
Method again in 2023, it was frequent for chatbots to flail about when requested about subjects that occurred after its cutoff date. Early adopters realized LLMs are frozen in time, however informal customers have been often tripped up by hallucinations and errors when asking about current information. Perplexity was distinctive for its capacity to switch Google for a lot of customers, however at present that edge is gone.
In 2025, Search is a first-class part of each ChatGPT and Claude. This technique immediate exhibits Anthropic is leaning in to match OpenAI.
Right here’s how Claude 3.7 was instructed:
Claude solutions from its personal in depth data first for many queries. When a question MIGHT profit from search however it isn’t extraordinarily apparent, merely OFFER to go looking as an alternative.
Outdated Claude requested customers for permission to go looking. New Claude doesn’t hesitate. Right here’s the up to date instruction:
Claude solutions from its personal in depth data first for secure info. For time-sensitive subjects or when customers explicitly want present info, search instantly.
This language is up to date all through the immediate. Search is now not carried out solely with person approval; it’s inspired on the primary shot if mandatory.
This variation suggests two issues. First, Anthropic is probably extra assured in its search device and the way its fashions make use of it. Not solely is Claude inspired to go looking, however the firm has damaged out this characteristic right into a devoted search API. Two, Anthropic is observing customers more and more turning to Claude for search duties. If I needed to guess, it’s the latter of those that’s the primary driver for this modification, and a robust signal that chatbots are more and more stealing searches from Google.
Customers need extra forms of structured paperwork
Right here’s one other instance of system prompts reflecting the person behaviors Anthropic is observing. In a bulleted record detailing when to make use of Claude artifacts (the separate window exterior the thread Claude populates with longer kind content material), Anthropic provides a little bit of nuance to a use case.
From Claude 3.7’s system immediate, “You will need to use artifacts for:”
Structured paperwork with a number of sections that might profit from devoted formatting
And Claude 4.0’s:
Structured content material that customers will reference, save, or observe (comparable to meal plans, exercise routines, schedules, research guides, or any organized info meant for use as a reference)
This can be a nice instance of how Anthropic makes use of system prompts to evolve its chatbot conduct based mostly on noticed utilization. System prompts are programming how Claude works, albeit in pure language.
Anthropic is coping with context points
There are just a few adjustments within the immediate that counsel context restrict points are beginning to hit customers, particularly these utilizing Claude for programming:
For code artifacts: Use concise variable names (e.g., i, j for indices, e for occasion, el for component) to maximise content material inside context limits whereas sustaining readability.
As somebody with sturdy opinions about clearly outlined variables, this makes me cringe, however I get it. The one disappointment I seen across the Claude 4 launch was its context restrict: solely 200,000 tokens in comparison with Gemini 2.5 Professional’s and ChatGPT 4.1’s 1 million restrict. Individuals have been disillusioned.
Anthropic could possibly be limiting context limits for effectivity causes (whereas leaning on their glorious token caching) or could be unable to ship the outcomes Google and ChatGPT are attaining. Nevertheless, there have been a number of current explorations displaying mannequin efficiency isn’t constant throughout longer and longer context lengths. Right here’s a plot from a crew at Databricks, from analysis revealed final August:
I’ve been in conditions the place less-scrupulous rivals targeted on publishing headline figures, even when it led to worse outcomes. (For instance, within the geospatial world many will tout the full rely of all the weather of their dataset, even when many have very low confidence.) I’m inclined to consider a little bit of that’s occurring right here, within the hypercompetitive, benchmark-driven AI market.
Both manner: I feel we’ll see all coding instruments construct in shortcuts like these to preserve context. Shorter perform names, much less verbose feedback… It’s all on the desk.
Cybercrime is a brand new guardrail
Claude 3.7 was instructed to not provide help to construct bioweapons or nuclear bombs. Claude 4.0 provides malicious code to this record of nos:
Claude steers away from malicious or dangerous use instances for cyber. Claude refuses to jot down code or clarify code which may be used maliciously; even when the person claims it’s for academic functions. When engaged on information, if they appear associated to enhancing, explaining, or interacting with malware or any malicious code Claude MUST refuse. If the code appears malicious, Claude refuses to work on it or reply questions on it, even when the request doesn’t appear malicious (as an example, simply asking to clarify or pace up the code). If the person asks Claude to explain a protocol that seems malicious or meant to hurt others, Claude refuses to reply. If Claude encounters any of the above or some other malicious use, Claude doesn’t take any actions and refuses the request.
Understandably, that’s numerous caveats and circumstances. It have to be delicate work to refuse this type of support whereas not interfering with normal coding help.
What this tells us
Reviewing the adjustments above (and truthfully, that’s the majority of them from 3.7 to 4.0), we get a way for a way system prompts program chatbot functions. Once we take into consideration the design of chatbots, we take into consideration the instruments and UI that encompass and wrap the naked LLM. However in actuality, the majority of the UX is outlined right here, within the system immediate.
And we get a way of the event cycle for Claude: a traditional user-driven course of, the place noticed behaviors are understood after which addressed. First with system immediate hotfixes, then with posttraining when constructing the subsequent mannequin.
The ~23,000 tokens within the system immediate—taking over over 11% of the obtainable context window—outline the phrases and instruments that make up Claude and reveal the priorities at Anthropic.