Sooner or later, humanity must learn to work with intelligence. CSF is an attempt.
For a long time, people have treated AI as a tool—striving to make it more obedient, more accurate, and more controllable. This direction is not wrong; it just does not address the most critical problem.
CSF approaches from a different direction: instead of taming AI, it establishes a collaborative structure with it.
CSF is the collaborative operating system for human-AI hybrid teams. Born from software engineering practice, with theoretical roots touching cognitive science, its core value lies in transforming AI from a tool into a reliable collaborator.
The starting point of CSF (Collaboration Specification Framework) is an axiom and two facts (which I call the Pang Principle—a small joke):
The Pang Principle (“Pang” is a Mandarin sound—fourth tone—representing the verbal shrug of “oh, of course we all already knew that.” It’s not really a “principle.” Call it a reminder. The name is a small joke—a nod to Huawei’s Tao Principle):
Natural language is inherently high-dimensional, probabilistic, and web-like, whereas symbolic systems are low-dimensional, deterministic, and linear. Attempting to bring certainty to a high-dimensional intelligent system using a low-dimensional symbolic system is like freezing water to keep it from flowing—you obtain certainty, but you lose vitality. Therefore, two facts demand our renewed attention:
The Pang Principle dictates the entire design direction of CSF.
Yet CSF is not a product of pure deduction. It has another side: the Owner put these preliminary ideas into practice during a real software product refactoring project (Bangzhao v3), letting the AI perform design work while iteratively refining the collaboration specifications. Through over 390 real sessions, the framework was continuously validated, corrected, and refined. Axioms were tested under pressure, specifications were rewritten in the wake of failures, and what ultimately emerged was a complete system spanning from theory to engineering implementation.
Logical self-consistency is its skeleton; real-world stress testing is its flesh and blood.
This origin endows CSF with a dual nature:
[!hint] The “Conflict” with Shannon’s Information Theory The axiom of the Pang Principle speaks of semantic information, whereas Shannon speaks of statistical information. These are two different dimensions of measurement:
- Shannon measures: How many bits are required to eliminate uncertainty.
- The Pang Principle measures: How much conceptual association, intent inference, and context completion an expression can activate.
The phrase “Help me do this well” has extremely low Shannon entropy (highly predictable), yet its semantic information volume is exceptionally high (requiring vast background knowledge to execute, or open to countless interpretations).
In one sentence: How humans and AI can reliably collaborate in long-term, complex, and goal-driven work.
Note the boundaries of this problem:
This problem holds true in the following three real-world scenarios, and CSF has been validated in all of them:
Scenario 1: Software Product Development (The Birthplace of CSF)
The Owner provides business goals and judgment; the AI Chief of Staff (CoS) handles architectural design and task planning; the AI Developer (Dev) executes the code. The three roles communicate via a file system, driving progress continuously across hundreds of sessions.
Scenario 2: Product Initiation and Requirements Reporting
A senior PM used CSF to write an initiation document for a data platform. He provided the business scenarios and requirement judgments, while the AI drove the structural progression. His description was: “It felt like collaborating with a thinking professional, to the point where I handed over the driver’s seat to the AI to some extent.”
Scenario 3: Commercial Bid Writing
A non-technical salesperson put the bidding documents and his own company materials into a directory, letting the AI draft the proposal according to the requirements. Beyond his expectations, the AI actively suggested bidding techniques he didn’t know and performed self-consistency checks. His evaluation: “It’s much easier than mentoring a human novice.”
The Common Structure of the Three Scenarios:
Human provides: Purpose + Business Knowledge (unique to humans)
AI provides: Structural Progression + Domain Expertise + Consistency Assurance
CSF provides: Collaboration Specifications that reliably bridge the two
[!NOTE] What exactly is CSF? How do you use it? CSF is essentially just a few Markdown files containing specifications written in plain, straightforward language. Using it is incredibly simple: drop these files into your workspace folder, open your AI chat window, and send: “Read context.” From there, the AI playing the “Chief of Staff” role in CSF will take over the progression, actively asking you questions and guiding you step-by-step through the discussion to get the job done. Throughout the process, the only thing you need to do is describe your purpose and thoughts—just like chatting with a colleague on Slack or Teams—and make choices or additions to the AI’s suggestions. You don’t need to understand code, nor do you need to learn complex prompts or configurations.
To understand CSF, one must first accept four engineering propositions. All other mechanisms are derived from here.
The first axiom of CSF is: A natural-language expression carries far more information than a symbolic system.
From this axiom, we deduce: The most effective means of controlling AI behavior is not exhaustive rules, persona settings, or formatting constraints—it is purpose.
A purpose is a single sentence, free of hidden assumptions, allowing the AI’s semantic system to automatically converge on the correct subset of capabilities. The more rules there are, the more brittle the system becomes; the clearer the purpose, the stronger it is.
Practical implication: All work units in CSF use purpose as their entry point. Once the purpose is defined, the methods and resources are determined accordingly.
From Proposition 1, we can infer that the nature of the collaborative relationship changes.
Humans are no longer the ones giving commands, and AI is no longer the one executing them. The driving mechanism shifts from “Go do X” to “The purpose is Y”—a shift that is not merely semantic; it redefines the roles of both parties in the collaboration:
Human holds: Right of Purpose Interpretation + Right of Business Judgment + Final Decision-Making Authority
AI holds: Right of Method Suggestion + Right of Knowledge Retrieval + Structural Progression Authority
This division of labor is not a polite agreement; it is an efficiency principle. The party that crosses the boundary is invariably replacing the other’s strongest capability with their own weakest.
When humans overreach to design methods, they replace the AI’s breadth of knowledge with empirical intuition. When AI overreaches to interpret purpose, it replaces human business judgment with probabilistic inference. Both forms of overreaching degrade the quality of collaboration.
This also leads to a redefinition of “errors”: we no longer expect AI to be error-free. Vitality is the source of intelligence, but it is also the source of uncertainty. Tolerance of error is not a lowering of standards; it is accepting the inherent cost of intelligence.
LLMs have two physical constraints: no cross-session memory, and a limited context window.
The mainstream approach is to “overcome” these constraints: external memory systems, expanding context windows, and using RAG to retrieve supplementary information.
CSF’s approach is to accept constraints and design within them:
Practical implication: CSF does not rely on any external technology stack (no RAG, no Agent frameworks, no memory systems). Technical complexity decreases as the design matures, rather than compounding as problems grow.
Mainstream assumption: Better results require better models (fine-tuning / RLHF / larger parameters).
CSF contends that because semantic dilution is inevitable, capability accumulation cannot be built on semantic quality; therefore, capability must be deposited in the engineering layer rather than the model layer.
Consequently, CSF advocates for an independent path in practice: instead of relying on model capability growth, achieve continuous improvement of systemic engineering capabilities by accumulating and organizing experience.
CSF spans three existing domains but does not fully belong to any single one.
The current AI engineering field already boasts numerous mature solutions: AI-assisted programming tools (Cursor, Copilot) address code generation quality; RAG solves knowledge retrieval and injection; Agent frameworks (LangChain, AutoGen, CrewAI) handle task automation; and Prompt Engineering optimizes single-turn dialogue output.
These solutions all address the same class of problem: how to make AI output better results.
CSF addresses a different problem: how humans and AI can reliably collaborate in long-term, complex work.
This is a different dimension. The former assumes humans are users and AI is a tool; CSF assumes humans and AI are collaborators, requiring a collaborative structure.
The practical corollary is that CSF is not mutually exclusive with the aforementioned solutions. You can use CSF within Cursor, or in a system equipped with RAG. CSF is a specification at the collaboration layer, not a replacement at the tool layer.
CSF draws upon the semantic boundary concepts of DDD (Domain-Driven Design) and the standardization approaches of classical software engineering, but introduces a fundamental modification:
Traditional methodologies assume that the executor possesses continuous memory. CSF must solve the problem of the executor (the AI) starting from scratch every single time.
A key proposition of CSF is: Many compromises of Agile methodologies were concessions to human labor costs in an era without AI. When AI can shoulder the burden of document maintenance, specification enforcement, and consistency checks, the rigor of classical software engineering becomes economically viable once again.
The design of CSF’s file system (three-tier indexing, lazy loading, baseline-log separation) shares similarities with knowledge management, but its purpose is different:
Knowledge management serves human retrieval and reflection. CSF’s external brain serves AI activation and localization—allowing a memoryless AI to immediately enter the correct working state upon every startup.
This is a concrete instantiation of the philosophical concept of Extended Cognition in AI engineering. It is akin to an Alzheimer’s patient using a notebook carried with them to record schedules and decisions.
AI Engineering: Solves single-turn dialogue quality, model capability, and tech stack setup
Software Engineering: Solves process specifications, team collaboration, and delivery quality
Knowledge Management: Solves information organization, retrieval, and personal productivity
The gap filled by CSF:
How human-AI hybrid teams can reliably collaborate in long-term, complex work.
CSF is the collaborative operating system for human-AI hybrid teams. Born from software engineering practice, with theoretical roots touching cognitive science, its core value lies in transforming AI from a tool into a reliable collaborator.