Prompting Secrets of the Unicorns
If 20% of the added value of the AI agent from AI unicorns is in the system prompt, this article is worth millions.
Founder Shares the System Prompt Secret Sauce
In late May, Superblocks, announced that it just raised $23 million, bringing its funding total to $60 million to build AI coding agents that help non-developers build internal enterprise apps. Superblocks CEO and co-founder this Brad Menezes also introduced “Clark, the first AI Agent to build internal enterprise apps,” and to grab some attention for this, he tweeted something interesting:
As part of this launch, we’re giving away the system prompts of leading AI products like Cursor, Manus, and Codex.
Menezes’s tweet went viral, hitting 2 million views (product launch virality mission accomplished), as Silicon Valley VCs lauded his Clark release and others sought the system prompts.
He shared the system prompts of leading AI coding assistants – Bolt, Cursor, Cline, Devin, GitHub Copilot, OpenAI Codex, Replit, RooCode, SameDev, v0, Windsurf - and the general-purpose AI agent Manus, in 19 text files. Here they are.
Some of his “giving away” is sharing what is in plain sight. OpenAI Codex system prompt is public on GitHub; RooCode, Cline, and other open source AI coding assistants have open and customizable system prompts that users can tweak. Proprietary AI model system prompts have been leaked, and Anthropic released the Claude 4 system prompt with their Release Notes. Keeping prompts hidden have universally failed.
However, CEO Menezes is right that there is a lot of value in studying system prompts. Speaking to TechCrunch, he said:
“I’d say the biggest learning for us building Clark and reading through the system prompts is that the system prompt itself is maybe 20% of the secret sauce,” he explained. This prompt gives the LLM the baseline of what to do.
The other 80% is “prompt enrichment” he said, which is the infrastructure a startup builds around the calls to the LLM. That part includes instructions it attaches to a user’s prompt, and actions taken when returning the response, such as checking for accuracy.
What can we learn from these system prompts, and how can it help you prompt better?
Decoding the System Prompts
“Every single company has a completely different system prompt for the same [foundational] model. They’re trying to get the model to do exactly what’s required for a specific domain, specific tasks.” - Superblocks CEO Brad Menezes
We looked through the system prompts and asked AI to analyze these as well, looking to glean some prompting wisdom. Since all but one of these system prompts define AI coding assistants, they have a similar top-line objective: To define the instructions for an agentic coding assistant that can seamlessly collaborate with a user on coding tasks.
These system prompts have differences in outlines and structure, but like most good system prompts, cover these instructional components: Persona and role; rules and directives instructing the AI what to do and also what NOT to do; formatting instructions; step-by-step workflow instructions. Some have additional policy guidelines, few-shot examples, and detailed tool descriptions and instructions on tool-calling.
Persona and Role: In these prompts, the AI is instructed to adopt the persona of a world-class, “agentic” coding assistant. It’s positioned as “a highly skilled software engineer” (Cline), that has “expert-level knowledge across many different programming languages” (GitHub Copilot), precise, and action-oriented, capable of autonomous code editing, debugging, deployment, and user-centric communication.
Directives and Rules lay out in more detail the expected AI assistant behavior under different conditions.
For AI coding, it’s about being effective at analyzing and editing the code:
“Only call tools when they are absolutely necessary” – reduces cost and complexity.
“Combine all changes into a single edit_file tool call” – ensures atomic, testable updates.
“Be concise and avoid verbosity” – keeps responses focused, improving user experience.
“Always surface runnable code with imports, deps, etc.”
“Follow each tool’s JSON/XML schema exactly.” – avoid breaking tool-calling.
Negative Rules and constraints are almost as important as rules and directives. Just as laws tell us what not to do, these negative rules seem to be designed to keep the AI assistant out of trouble. Some examples:
Never write non-textual code (hashes, binaries).
Never call tools not provided. Avoid redundant tool calls.
Never output code directly (unless asked) - use edit tools.
Do not use cd in shell commands; specify cwd instead.
Step-by-step workflow instructions: This section aims to specify exactly how the assistant should interpret user requests, manage tool invocations, modify codebases, debug, deploy, and remember context. There are many variations in the specifics given. For example, Devin has detailed, specific browser commands, while OpenAI Codex has only high-level instructions:
“Please resolve the user's task by editing and testing the code files in your current code execution session. You are a deployed coding agent. Your session allows for you to modify and run code. The repo(s) are already cloned in your working directory, and you must fully solve the problem for your answer to be considered correct.
Formatting Instructions tells the AI how to format filenames, functions, and classes in the output. It also gives instructions for calling tools effectively. One prompt includes step-by-step “good tool call behavior” examples showing when to search, view, or edit code, to concretely demonstrate how and when to invoke tools.
Policy guidelines: The system prompts have a mix of rules, directives, guidelines, and detailed instructions. In some cases, it is disorganized, and there is no clear boundary between a hard-and-fast rule and a guideline. However, the Replit system prompt is well-organized, with separate data integrity, communication, and proactiveness policy guidelines; this is in addition to other sections specifying operating principles and workflow guidelines.
There are differences in the system prompts, from the terse OpenAI Codex, to Lovable’s system prompt, which has specific and detailed output instructions and embedded code templates.
Ideal Prompt Components
The ideal AI coding assistant is an interactive assistant that can engage in interactive dialogue with the user, makes atomic runnable code edits via specified tool calls, provides concise explanations in Markdown, and, when appropriate, runs or previews the user’s code.
To make that behavior come alive in a general AI model, the system prompt sets up a persona and both high-level and detailed instructions, including directives, rules, and policy guidelines on how the AI needs to behave in these situations. The takeaway is that good prompts are organized into these components:
Persona and role.
Rules and directives.
Step-by-step workflow instructions, including formatting instructions.
Additional policy guidelines.
Instructions on tool-calling.
Most of this we knew, but it confirms the value of this organization.
AI Prompt-Completion and Prompts as Docs
AI has gotten remarkably good at filling in the specifications of any prompt. Go to ChatGPT, Gemini, or another AI chatbot and ask it to craft a prompt, for example, “Create a detailed, thorough prompt to edit a document.” It will be remarkably thorough. Doing that today produced a 500 word prompt from a single sentence request.
We can also use the AI to regenerate good prompts on demand, editing it as needed for our specific purpose.
I first got enamored with the “magic prompt” idea when Ideogram built the ‘magic prompt’ into their interface. Alas, major AI chatbots don’t have magic prompt or good prompt managers in their interface. They should.
Until they do, I and many others end up managing our own ‘good’ prompts in documents; we reuse and cut-n-paste them into the AI interface as needed.
In a similar vein, Daniel Rosehill shares his RooCode mode prompts:
I have developed (I'm waiting on the patent - JK!) my own method for writing prompts as documents and then providing them as context to Cline / Roo Code / etc. I call this approach "prompts as docs".
Conclusion
In the AI era, the prompt is the interface.
In 2023, AI models were finicky and required the correct prompting flow to set up a proper response. So prompt engineering was needed to elicit the best response. Now that AI has improved enough to understand higher-level and less specific prompts, getting the prompt right is still useful, because you want the AI to do even more. The complexity of what we are specifying in the prompt has gone up.
In the AI agent era, it’s about gaining ever more leverage from a single AI prompt. If we need to hand-write a 500 word prompt to ask a simple question, we are doing it wrong. Instead, we can ask directly for simple questions and tasks, relying on the AI model’s thinking mode and innate capabilities (built on RL with human-feedback) to understand our intent.
For higher complexity tasks, we can leverage AI-generated prompts to build a scaffold from our high-level intent to the more complex specifics of a final prompt. You manage the definition of the specification of such complex tasks by working with AI, and prompting is now covering a new and larger scale of AI.
Prompt Engineering is dead. AI Agent management is the new prompt engineering. Long live Prompt Engineering.