Tacit Knowledge: The (Hopefully) Irreplaceable Skill in the AI Era

With AI agents becoming increasingly prevalent, a rather eerie trend has surfaced on the internet, sounding much like an episode of Black Mirror.

Someone released a program on GitHub called “Colleague.skill” [1]. It creates a digital avatar for a colleague who has left the company: by uploading their Lark messages, DingTalk documents, WeChat chat logs, emails, screenshots, and anything else bearing their digital footprint, along with your description of them, the program uses these materials to generate an AI Skill. You can then build an agent based on this to do their work.

The colleague has resigned, and their desk is empty. Yet, the agent representing them is diligently working overtime in the company, perhaps even more industrious than before.

Isn’t this tragic? People inevitably wonder if they will be next. Even worse, next time, will the company extract my skill first and then lay me off directly?

This discussion might offer you some reassurance. I do not believe your value to the company can truly be distilled into a SKILL.md document. Even if AI can mimic your tone and replicate your workflows, it still cannot thoroughly replace you—because it cannot take away your core asset, which is known as “Tacit Knowledge”.

What is Tacit Knowledge? #

Simply put, tacit knowledge refers to knowledge that “can be understood but not described.” You know how to do something, but you cannot explain it clearly, nor can you write a manual to teach it to others. Its counterpart, which can be articulated, is called “Explicit Knowledge”.

For instance, if you are learning to cook and are given a recipe that says: “Heat oil over medium heat, stir-fry scallions and ginger until fragrant, reduce the sauce until it coats the spoon.”

You recognize every word, yet you still don’t know how to cook the dish. What exactly is “medium heat”? It is not just the knob on the gas stove; it depends on the wok, the burner, the amount of oil, and the water content of the ingredients. Is “fragrant” a chemical reaction between scallions, ginger, and oil occurring at the 12th second? That is a critical moment that must be judged by sound, smell, and color simultaneously. And what does “coat the spoon” even mean?

You need a master chef on-site to teach you. Even he cannot clearly articulate the specific criteria, but he can point them out for you: “You can flip it now.” “Wait another second.” “Smell that? This is the exact moment.”

You must feel it with your own heart.

What is written in the recipe is explicit knowledge; the skills that cannot be written down are tacit knowledge. In the words of British analytic philosopher Gilbert Ryle, the former is “knowledge-that”, and the latter is “knowledge-how” [2]. The former involves Standard Operating Procedures (SOPs), focusing on definitions, steps, and formulas; the latter focuses on feel, timing, discretion, and rhythm.

Explicit knowledge is sometimes kept secret, but tacit knowledge requires no secrecy. The chef tosses the wok right beside you, the surgeon operates right in front of you, the senior programmer debugs while sharing their screen. You see the entire process, yet you still cannot learn it.

If all your knowledge comes from text-based explicit knowledge, it is extremely dangerous to jump straight into hands-on practice. Practical application demands tacit knowledge.

The Ineffable Process of Internalization #

The concept of tacit knowledge was first proposed in 1958 by Michael Polanyi, a Hungarian-British philosopher [3].

Polanyi had a signature quote: “We can know more than we can tell.” He compared knowledge to an iceberg floating in the ocean: the small tip exposed above the surface is explicit knowledge, while the massive body submerged underwater is tacit knowledge.

Before becoming a philosopher, Polanyi was a physical chemist. Early on in the laboratory, he realized that what truly enables a scientist to make judgments is not just the published methods in papers, but a vast, unspoken set of intuitions, feel, experiences, and beliefs.

Why is tacit knowledge difficult to articulate? Polanyi explained that it is a form of “Integration”: we do not view things in isolation, but rather synthesize many subtle clues toward an overall meaning.

For example, when you look at a familiar face, you don’t measure eyebrow length, nose angle, and eye distance ratio before calculating, “This is Zhang San.” You recognize them instantly. A pianist doesn’t stare at each finger and piece them together into music; quite the opposite, they have to “forget” the fingers for the music to emerge.

It’s like hitting a nail with a hammer. Your attention is on the head of the nail, but your hand muscles, your grip strength, and the angle of your arm are constantly making highly complex, real-time corrections. These minute adjustments are not only indescribable, but they should not be described. If you try to break down those movements into language, you won’t be able to hammer the nail.

Though ineffable, those background details—which constitute tacit knowledge—are precisely the foundation of your understanding and performance.

To master tacit knowledge, you must internalize those details. In Polanyi’s terms, this is called “indwelling” [4]. First, you dwell within the tools, skills, and details; eventually, they grow into you, and you become one. Just as a blind person’s cane is no longer merely a stick in their hand but an extension of their body and perception. You don’t know how to ride a bicycle because you memorized the physical instructions; you can ride it because the bicycle has, in part, dwelled within your body.

In modern terms, mastering tacit knowledge means updating and fine-tuning the parameters of your brain’s neural networks. You can’t explain exactly which parameters changed or how, but they undoubtedly did.

Four Key Characteristics of Tacit Knowledge #

Tacit knowledge differs fundamentally from explicit knowledge in four ways, making it resistant to being distilled into a SKILL.md file.

First, tacit knowledge involves complex pattern recognition. A senior doctor can intuitively diagnose a rare disease just by looking at a patient’s complexion. How did they see it? Which specific pixel told them? It’s impossible to say. This intuition is the internalization of long-term practice—a high-dimensional compression of ten thousand trials and errors.

Second, tacit knowledge is highly context-dependent. You can write a litany of rules stating what to do in situation A and situation B, but the real world always presents exceptions. A top-tier salesperson employs different communication strategies depending on culture, mood, or even weather, demonstrating a “sense of propriety” that adapts to the environment. You simply cannot write out all those variations.

A novice doctor relies on guidelines, while a master doctor evaluates whether “this patient is currently suitable for following the guidelines.” A duty manager doesn’t just read a flowchart; they judge whether an anomaly is truly an anomaly.

Third, tacit knowledge is often a collective property—knowledge that exists between people [5]. When a team stays together long enough, they develop an unspoken consensus: when to “hold your tongue,” when to press further, what crosses the line, what kind of joke is acceptable, which stance requires caution, what it means for a statement to be “technically correct but politically incorrect,” and when it’s best to cancel today’s meeting. These are not individual inner monologues; they are the tacit rapport forged over time by the entire community.

Fourth, tacit knowledge often requires bodily participation. In cognitive science, the theory of “Embodied Cognition” [6] suggests that much knowledge resides not only in your brain but also in your muscle fibers and nerve endings. Timing, feel, balance, tone, rhythm, pauses, and the exact moment to look up—these are millisecond-level integrations. They cannot be fully captured by the linear bandwidth of natural language.

How do you distill this knowledge? A human must be present in the field. SKILL.md can only migrate an employee’s explicit knowledge; it cannot restore their soul.

How to Learn Tacit Knowledge #

How, then, do we learn tacit knowledge? Not to worry. As long as it is knowledge, it can be transmitted. The issue is that it cannot be taught through manuals alone—you must be present in the field to dwell within it.

Today, despite the vast reach of the internet, professions like surgeons, pilots, and even carpenters still retain the oldest knowledge transmission system: apprenticeship. You need a master by your side.

First, you observe on-site. You cannot expect to learn surgery by watching videos. You must be in the operating room, standing next to the master, observing how they hold the scalpel, feeling the calm atmosphere when they handle sudden massive bleeding, and seeing how they coordinate with the nurses.

Next, you do it alongside them. The master demonstrates, and you imitate. They provide real-time feedback and corrections. If you miss something, they bring background details to the foreground for closer inspection. You engage in deliberate practice, piece by piece, until these details slowly infuse your body and form a bodily schema.

If you don’t have a formal mentor, observe the experts around you who handle complex interpersonal issues or crises. Notice how they communicate outside formal meetings, their micro-expressions, and their sense of rhythm. Then, figure it out yourself through practice.

Will AI Devour Tacit Knowledge? #

Here comes the most crucial question: Is it possible for AI to devour human tacit knowledge?

Polanyi passed away in 1976; he didn’t live to see the AI we have today. His favorite example—facial recognition—has now been easily conquered by AI. Robots can learn to ride bicycles [7], and bicycles can even balance themselves. Machines do not need you to write out explicit recognition rules; through massive data and training, they silently absorb countless details and grasp the very things that rules cannot articulate.

Ultimately, AI is a neural network, and isn’t the human brain also a neural network?

If text records cannot represent the entirety of a colleague, what if I combine them with video records? What if I document every movement, every spoken word, and every micro-expression? The operating room has already become a data-native environment, where multimodal information streams support real-time decision-making, predictive modeling, and personalized workflows [8]. Why couldn’t we train a robot surgeon?

The reality is that AI is already nibbling at the outer edges of tacit knowledge. AI diagnostic accuracy has surpassed that of human doctors. Physicians are increasingly accepting—and even relying on—AI diagnoses. We are witnessing AI employees outperforming human employees.

“Ineffable” does not mean “untrainable for AI.” “Cannot be written into SKILL.md” does not mean “cannot be integrated into large model parameters.”

However, the reflections of our predecessors on tacit knowledge still give us some confidence. Perhaps AI will ultimately fail to exhaust all tacit knowledge. We have at least three reasons for this:

First, drawing from Polanyi again: he believed completely explicit knowledge is inconceivable because the application of any explicit rule must be grounded in a deeper, unutterable background of tacit internalization. Behind a theorem lies a deeper theorem, ultimately tracing back to a few axioms. And behind those axioms, there are no other axioms; they are simply unreasonable—they are our embodiment and faith. Will AI dare to make leaps of faith?

Second, from sociologist of science Harry Collins, who proposed the aforementioned “collective tacit knowledge” [5]. His insight is that human interaction is a dynamically evolving process. Even if a machine could perfectly capture yesterday’s statistical patterns, it has no way to predict the new tacit rapport we will develop tomorrow. Xiao Li has left; our interaction with his digital avatar could never be identical to our interaction with the real Xiao Li today. Perhaps a company’s true competitiveness stems precisely from that difference.

Third, embodied cognition. Why is it that AI is already better than humans at playing chess and solving math problems, but getting a robot to clear a dining table is so difficult? This is known as “Moravec’s paradox.” The answer may be that the tacit knowledge required to interact with the physical world is the result of hundreds of millions of years of human biological evolution. It might not be easily digitized.

In fairness, I do not believe these reasons are absolute absolutes. But they are highly valuable points for reflection. For the time being at least, if humans still hold any advantage, this tacit knowledge is our moat.

Building Your Moat in the Age of AI #

So, in the age of AI, how should we build our competitiveness on tacit knowledge? The intuitive thought might be to avoid digitizing my craft, preventing people from seeing how I work. I will only present results without explaining the steps to avoid being distilled into a skill. But this approach is a dead end.

Rest assured, as long as “turning humans into agents” is profitable, a peer with better skills will eventually sell their craft to an agent company. Instead of hiding, you are better off writing down all the articulate-able workflows to create leverage and boost your efficiency.

If yesterday’s tacit knowledge is destined to be automated, we must invest our energy in tomorrow’s new tacit knowledge.

First, this means claiming high-context, high-exception, high-responsibility positions. These roles involve frequent anomalies, rapid feedback, ambiguous boundaries, blurred responsibilities, and frequent value conflicts. Operations here are hard to standardize. Diagnosing, negotiating, architecting, curating topics, editing, hiring, crisis management, brand judgment… Every micro-decision will sediment into your new intuitive feel.

If you can’t seize such a position immediately, you must still enter the field to “dwell within.” You absolutely cannot rely entirely on theory or let AI do everything for you. You aren’t just doing this to get the job done right; you are doing it to accumulate compound interest. You have to grow the skills onto yourself. What AI knows is public infrastructure; tacit knowledge is your private capital. Otherwise, why would you be offered a more important position in the future?

In the long run, the most valuable tacit knowledge anyone can accumulate is their unique aesthetics, persona, and trust relationships.

Your linguistic rhythm, your judgment style and sense of humor, your value function, and your tacit rapport with the crowd—even if AI can imitate them 100%, it’s useless. Because what people want to see is the dynamic you; they want to know what surprises you will bring next. Someone might successfully use AI to write today’s commentary in the style of Lu Xun, but that isn’t what we want.

What we want is: if Lu Xun were alive today, experiencing things firsthand and witnessing these new chapters of history, how would he evaluate current events? We want to invoke the real Lu Xun, not his digital avatar.

Ultimately, on the day when material abundance reaches its peak, what people will care about in each other is certainly not their skill tree, but the person themselves.

The Wheelwright’s Revelation #

Speaking of tacit knowledge, we cannot ignore a short story from the Zhuangzi called “Wheelwright Bian.”

Duke Huan of Qi was reading a book in the hall, and a wheelwright named Bian was carving a wheel down below. Bian asked Duke Huan, “What book are you reading?” The Duke replied, “The words of the sages.” Bian sneered and said, “Those are the dregs of the ancients!”

Duke Huan was furious and declared that unless Bian could provide a good reason, he would put him to death. Bian gave an explanation that even Polanyi, two thousand years later, could not surpass:

“Let’s take making wheels as an example. If I work too slowly, it doesn’t work; if I work too fast, it doesn’t work either. Exactly how to apply the force—I cannot express it in words, and I can’t even teach it to my son. That’s why I am seventy years old and still here working as a carpenter. So tell me, how could the craft of the ancients all be written down in a book? Aren’t the explicit knowledge you are reading just dregs?”

If you hear that your former company has turned you into a SKILL.md, you can rightfully tell them: What you have distilled are merely my dregs.

【Closing Poem】

What can be spoken are the rules, What can be executed is the mastery. Models can memorize your sentences, But discretion only grows on the paths you have walked.

References #

1. titanwings/colleague-skill
2. Ryle, Gilbert. “Knowing How and Knowing That: The Presidential Address.” Proceedings of the Aristotelian Society 46 (1946): 1–16.
3. Polanyi, Michael. Personal Knowledge: Towards a Post-Critical Philosophy. Chicago: University of Chicago Press, 1958.
4. Hadjimichael, Demetris, Rodrigo Ribeiro, and Haridimos Tsoukas. “How Does Embodiment Enable the Acquisition of Tacit Knowledge in Organizations? From Polanyi to Merleau-Ponty.” Organization Studies 45, no. 4 (2024).
5. Collins, Harry. Tacit and Explicit Knowledge. Chicago: University of Chicago Press, 2010.
6. Shapiro, Lawrence. “Embodied Cognition.” The Stanford Encyclopedia of Philosophy, 2021.
7. Liu, Jinghao, et al. “Coordinated Trajectory Tracking and Self-Balancing Control for Unmanned Bicycle Robot Against Disturbances.” Actuators 15, no. 1 (2026): 49.
8. Kudsi, Omar Y., et al. “Reimagining Surgery in a World Powered by Intelligence, Connectivity, and Code.” npj Digital Surgery (2026).