---
title: "Skills that cannot be written down"
description: "Polanyi-style tacit skill, why prompts can't capture it, and how crystallization can."
doc_version: "1.1"
last_updated: "2026-05-01"
canonical: "https://modiqo.ai/blog/skills-that-cannot-be-written-down.txt"
---

# Skills That Cannot Be Written Down

*On Michael Polanyi, the blind man's stick, the swimmer who doesn't know he holds his breath, and why Rote's skills look nothing like the skills in your README.*

---

## The word "skill" is doing two different jobs

If you have worked with modern agent frameworks, you have encountered "skills." They are usually markdown files. A skill might say: *When the user asks about invoices, query the billing database using this schema, format the result in this table, and return.* It is a document. It is instructional. It is essentially a small runbook the agent consults before acting.

This is useful. Instructional skills have a real place and a real necessity. They encode what's already been articulated — conventions, schemas, required fields, known procedures. They are the organization's operating manual handed to the agent. Nobody should be against them.

But when we started building Rote, we kept running into a different kind of skill that couldn't be captured that way. The skill of *figuring out which endpoint to probe when the obvious one returns stale data.* The skill of *recognizing, three calls in, that this API is pretending to support pagination but actually doesn't.* The skill of *knowing, after years of dealing with a particular payment processor, that certain errors mean retry and certain errors mean stop, even though the status codes are the same.*

These are skills too. Senior engineers have them. They cannot hand them to a junior on day one as a document. They can barely articulate them when pressed. And yet they are the difference between an expert operating in a system and a novice fumbling in it.

We did not have the vocabulary for this category of skill. Until we read Michael Polanyi.

---

## Who Polanyi was, briefly

Michael Polanyi was a Hungarian-British polymath — a physical chemist of the first rank who made original contributions to reaction kinetics and crystal structure at Berlin and Manchester before turning, in middle age, to philosophy. His magnum opus, *Personal Knowledge: Towards a Post-Critical Philosophy*, came out of his 1951–52 Gifford Lectures and was published in 1958. His later, shorter *The Tacit Dimension* (1966) was based on the 1962 Terry Lectures at Yale. Between these two books he laid out one of the most important epistemological frameworks of the twentieth century — a framework that quietly dismantled the positivist dream of reducing all knowledge to formal propositions.

The single sentence that carries the most weight across his work is this one, from *The Tacit Dimension*: *"We can know more than we can tell."*

Polanyi meant this literally. Not as a humble disclaimer but as a structural claim about the architecture of knowing. There is knowledge you possess and use reliably that you cannot write down. Not because you are lazy, not because the medium is inadequate, but because the knowledge itself has properties that resist articulation.

Two ideas from Polanyi reshaped how we think about what Rote is doing. The first is from Chapter 4 of *Personal Knowledge*, section 22, titled "Tools and Frameworks." The second is from what Polanyi called the "tacit component" of knowing. Each corresponds, strikingly cleanly, to one of Rote's core primitives.

---

## Part I — Tools, probing, and the extension of the body

Polanyi asks you to do something simple. Pick up a stick. Touch it to an object. What do you feel?

The first time a blind man picks up a probe, Polanyi writes, his fingers register a series of disconnected shocks. The stick hits something, transmits a jolt, hits something else. There is no coherent world at the tip — just noise at the hand. But something remarkable happens with practice. The jolts in the fingers gradually cease to be felt as jolts. Instead, they become *the presence of obstacles of certain hardness and shape, placed at a certain distance, at the point of the stick.* The stick is no longer an object being held. It has become an extension of the body. The man attends *from* his fingers *to* the world the probe discloses.

Polanyi gives this structure a name. He calls it *subsidiary awareness* of the tool, integrated into *focal awareness* of what the tool reveals. The sensations in the fingers have not gone away. They have been *assimilated* — they are still there, but the knower is no longer attending to them as such. He is attending through them, to something else.

This is, Polanyi insists, the general structure of tool use. A hammer becomes part of the hand that swings it. A surgical probe becomes part of the surgeon exploring the cavity. A motor car, a tennis racket, a hammer — each one, once mastered, passes from the category of *object* into the category of *extension of self*. *"We pour ourselves out into them,"* he writes, *"and assimilate them as parts of our own existence. We accept them existentially by dwelling in them."*

The word *dwelling* is doing a lot of work here. It is not a metaphor. Polanyi is claiming that the knower literally comes to inhabit the tool — that the tool becomes part of the body through which the world is known. And — this is the key move — *the same structure applies to intellectual tools*. Frameworks, theories, interpretive vocabularies. The scientist does not stand outside quantum mechanics and manipulate it; the scientist dwells in it and perceives the world through it. *"When we accept a certain set of pre-suppositions and use them as our interpretative framework, we may be said to dwell in them as we do in our own body."*

Now read that paragraph again with an agent in mind.

---

## What this clarified for us about adapters

The standard way of giving an LLM access to an external system is to hand it a tool specification. The tool has a name, a description, a set of parameters, and a return type. The agent calls the tool and receives a response. This is functional. It works. It also completely misses what Polanyi identifies as the *point* of a tool.

A tool, in Polanyi's sense, is not the interface specification. The tool is the thing the agent probes the world with — and the signal that actually matters is *what the agent learns on the far side of the probe*. When a senior engineer uses a REST API for the first time, she doesn't read the OpenAPI spec and then "know the API." She calls an endpoint. The response has a field that wasn't documented. She notes it. She calls another endpoint. The pagination is off by one. She notes it. She tries a combination the docs said would work and it doesn't. Somewhere around call fifteen, she has a *feel* for the API. She knows, tacitly, which endpoints are reliable, which responses are lying, which edge cases matter. That feel is the skill. The spec was the starting point, not the skill itself.

This is exactly the blind man's probe. The stick is useless until you've hit things with it. The feel is the thing that cannot be written down.

This shaped how we thought about **adapters** in Rote. An adapter is not a specification document that describes an external system. An adapter is the surface the agent probes the world through — and critically, it is *instrumented to capture what the agent learns on the other side of each probe*. The adapter's job is not only to let the agent call the tool. Its job is to turn the return values of those calls into something that can be accumulated, examined, and eventually turned into a feel for the system.

This reframes what "search" means for an agent. An agent solving a problem against an unfamiliar system is doing something Polanyi would have recognized instantly: it is probing, integrating the shocks into a coherent perception, gradually making the tool *disappear* so that what comes into focal awareness is the system on the other side. The probe-call loop is not a brute search over an API graph. It is the structural process by which the agent's subsidiary awareness of individual call results gets integrated into focal awareness of a *system* — a shape, a topology, a set of reliable affordances.

When we talk internally about "search space abstraction" in Rote, this is what we mean. The space is not the API. The space is the topology the agent discovers by probing the API. The abstraction is not the spec. The abstraction is the feel.

Polanyi, in 1958, gave us the vocabulary we were missing. Every tool the agent uses starts as an external object — a list of endpoints, a schema, a wall of documentation. Every tool the agent *masters* becomes an extension of the agent, something dwelt in rather than read. The adapter is the mechanism by which this transition is made possible. Without that structure, the agent is perpetually a first-day engineer holding a stick and feeling disconnected shocks.

---

## Part II — The tacit component, and why instruction alone cannot carry a skill

Set down *Personal Knowledge* for a moment. Pick up *The Tacit Dimension*. Polanyi opens with a claim that sounds modest and turns out to be radical: *we can know more than we can tell.*

He illustrates it with the simplest example he can find. A skilled swimmer, Polanyi observes, keeps himself afloat by regulating his respiration — specifically by never fully exhaling, so that the lungs remain partially inflated and buoyancy is maintained. This is an exact, technical fact about how swimming works. And yet almost no swimmer, when asked, can tell you this is what they are doing. They learned to swim. They stopped sinking. They did not, at any point, think "I must keep my lungs at approximately sixty percent inflation to maintain neutral buoyancy." They felt their way to it. The knowledge was acquired in the body and remained there — unspecifiable, irrecoverable as words, but perfectly reliable in action. The swimmer knows more than the swimmer can tell.

This is Polanyi's point about skill, about expertise, about the kind of knowing that actually makes practitioners effective. Most of it cannot be written down. And the parts that *can* be written down are usually not the parts that matter.

So how does skill actually transfer from one person to another, given this? Polanyi's answer is the oldest answer in the world, and the one our culture keeps trying to abolish. It transfers through **apprenticeship**. Observation. Imitation. Reflection. Correction in context, across many trials. *"By watching the master and emulating his efforts in the presence of his example,"* Polanyi writes, *"the apprentice unconsciously picks up the rules of the art, including those which are not explicitly known to the master himself."*

This last clause is the one to sit with. The apprentice picks up rules the master *doesn't know he's following*. The transfer works not because the master explains, but because the apprentice *reflects on the master's doing, while doing alongside*. The knowledge was never articulate. It was never going to be articulate. It is transferred body to body, practice to practice, across a period of shared attention.

---

## See one, do one, teach one

The surgical tradition has a name for this: *see one, do one, teach one.* The phrase is attributed to William Stewart Halsted, who became the first Chief of Surgery at Johns Hopkins Hospital in 1890 and invented the modern surgical residency. Before Halsted, American surgical training was informal — most surgeons were effectively self-taught. Halsted's innovation was to insist that surgery could only be learned the way Polanyi later described knowledge being learned in general: by doing, in proximity, across many cases, under the correction of someone who had already done it.

A century and a half later, in an age of simulators and AI-augmented training, the phrase still circulates in operating rooms. Not because surgeons are traditionalists but because the phenomenon is real. You cannot teach a resident to suture by handing them a textbook. You cannot teach them to feel the difference between healthy tissue and compromised tissue by writing it down. They have to *see* it, *do* it alongside someone who has done it enough times that they no longer think about it, and eventually teach it themselves — at which point they discover how much of what they now do automatically is invisible even to them. The cycle is the mechanism by which tacit knowledge is generated, extended, and passed on.

There is a charming industrial version of the same story. In 1985, Matsushita (now Panasonic) was trying to build an automatic home bread-making machine. The team couldn't mechanize the dough-kneading process. Their loaves were wrong. Every bread recipe in every cookbook produced a machine that made bad bread. So a software engineer named Ikuko Tanaka went to apprentice with the head baker of the Osaka International Hotel, who was reputed to make the best bread in the region. She watched. She kneaded dough alongside him. And eventually she observed something the baker had never mentioned and never thought to mention — that he was not only *stretching* the dough but also *twisting* it in a particular motion, a "twisting stretch" that turned out to be essential to the texture. The baker had never said this, in the same way a swimmer never says they keep their lungs inflated. He did not know he knew it. His apprentice discovered it by watching him act, and only then could it be fed back into the machine.

This is what tacit knowledge transfer actually looks like. It is not a transcription. It is a reflection-on-doing, across cases, until a pattern surfaces that neither party could have articulated in advance.

---

## What this clarified for us about the canvas

Modern agent systems do the opposite of apprenticeship. They instruct. The skill file says: *do it this way*. The system prompt says: *follow these steps*. The agent reads the instruction, tries to follow it, fails in some case the instruction didn't anticipate, and either asks for help or hallucinates a solution.

This is the pedagogical equivalent of handing a medical student a textbook and asking them to perform surgery. It is using the one modality — explicit propositional transfer — that Polanyi specifically warned us has the lowest signal-to-noise ratio for how skill actually moves between practitioners. The instruction can carry the easy parts. It cannot carry the feel.

Which raises the question: what would it look like to build an agent architecture that learns *the way a swimmer learns*, or *the way an apprentice learns*, rather than the way a student reads a textbook?

Rote's **canvas** is our attempt at an answer. The canvas is the workspace in which an agent does the task. It is not the instruction; it is the place where the agent's actions leave traces. Every probe, every call, every decision, every correction is visible on the canvas. The agent acts. The canvas records what the agent did and what happened as a result. Over time, the record accumulates into something that looks less like a transcript and more like a portfolio of *practice*.

The point of the canvas is not observability for its own sake — though observability is a happy byproduct. The point is that the canvas is the surface on which an agent can *reflect on its own doing*. When the agent solves the same kind of problem five times on the canvas, the canvas contains five instances of the agent feeling its way forward. Some steps are the same across all five. Some are specific to one case. Some are wrong on three of them and right on two. The canvas holds this. And because it holds this, it can be *examined* — by the agent itself, by another agent, or by a human — for the structural regularities that run through the cases.

This is the machinery of apprenticeship, rendered in software. The agent is not being told how to do the task. The agent is doing the task, many times, in a place where its doing is preserved. What it learns is not what it was told. It is what surfaced through the doing.

Polanyi's swimmer never knew she held her breath. Ikuko Tanaka never would have known about the twisting stretch if she had only read the baker's recipe. The canvas is how we give an agent the same shape of experience — a place to swim, a kitchen to knead dough in, something to feel its way through.

---

## From canvas to flow: crystallizing tacit knowledge

Something important happens after the apprentice has done the same thing enough times. The doing stops being deliberate. It becomes — Polanyi uses this word repeatedly — *assimilated*. The swimmer no longer thinks about breath. The surgeon no longer thinks about suture tension. The motion is no longer composed out of attention to its parts. It is executed as a whole, while attention goes elsewhere.

This is the moment at which tacit knowledge has *crystallized*. It exists. It operates. It is deeply reliable. And it has passed out of the domain of deliberate thought.

Polanyi emphasizes that this transition is *irreversible*. You cannot un-learn to swim. You cannot un-integrate the stick you have been using for years. The knowledge has become part of you. And — here is the critical consequence for our purposes — once knowledge has crystallized, it can be *transferred*. Not as instruction, because it was never instruction. But as *demonstration*, as *performance*, as something the next apprentice can do alongside you until they too have assimilated it.

Rote's **flows** are crystallized canvases. When the agent has probed the adapter enough times, when the canvas has accumulated enough practice of the same task, the structural regularity of what actually worked across cases can be extracted and saved as a flow. A flow is not a set of instructions. It is a deterministic replay of the part of the exploration that turned out to matter, stripped of the parts that didn't. It is the twisting stretch, once someone has watched the baker long enough to see it.

Crucially, a flow is *transferable tacit knowledge*. It is the crystallized form — which means it can be invoked by another agent who has never seen the underlying system, the same way a junior engineer can follow a senior engineer's established procedure even though the junior has never personally acquired the feel behind it. The flow carries what the canvas discovered. The next agent does not need to rediscover it.

This is the inversion that matters. Instructional skills say: *here is what to do*. They assume the knowledge can be articulated and handed over. They work for the easy parts. Rote's skills — the ones that emerge from canvases and crystallize into flows — say: *here is what was done, across many cases, where it mattered*. They do not assume articulation. They preserve the practice. They transfer the feel, because the feel has been compressed into structure that can run the same way again.

Instructional skills treat the agent as a student reading a manual. Rote's skills treat the agent as an apprentice who learned by doing, in a workspace that remembered what the doing revealed.

---

## Why this matters more than it sounds

If you have stayed with me this far, the frame should feel less abstract. Two of Polanyi's claims map onto two of Rote's primitives almost perfectly:

The tool-and-framework structure of *Personal Knowledge* — subsidiary awareness integrated into focal awareness, the probe that becomes an extension of the body — is what adapters are trying to capture. An adapter is not a spec. It is the surface the agent probes the world through, instrumented so that what the agent learns on the far side can accumulate.

The tacit-knowledge structure of both books — knowledge acquired by doing, preserved in the body of the practitioner, transferable only through reflection on shared practice — is what the canvas-and-flow pair is trying to capture. The canvas is the apprentice's workspace. The flow is the crystallization.

The opposite of this architecture is the instructional skill: the markdown file that says *here is how to do the task*, handed to the agent as a precondition. Instructional skills have their place. They are excellent for the parts of the work that genuinely are articulate. But they are trying to carry, with the wrong vehicle, the kind of knowledge that Polanyi showed us sixty-eight years ago cannot be carried that way. A swimmer cannot be taught to swim by being told about buoyancy. A baker cannot teach the twisting stretch by writing it in the recipe. A senior engineer cannot hand her API intuitions to a junior by updating the README.

The reason agents have been so brittle, so expensive, so prone to burning tokens rediscovering what they already did yesterday, is that the industry has been trying to build agents the way a terrible teacher teaches — by handing them more and more instruction, and expecting instruction to carry what only practice can carry.

Polanyi saw this with perfect clarity in the middle of the last century. *"Tacit assent and intellectual passions, the sharing of an idiom and of a cultural heritage, affiliation to a like-minded community: such are the impulses which shape our vision of the nature of things on which we rely for our mastery of things. No intelligence, however critical or original, can operate outside such a fiduciary framework."* Mastery is not transferred by rules. It is transferred by dwelling. Intelligence cannot operate outside the framework it has come to inhabit.

The whole architecture of Rote is an attempt to take this seriously. Adapters so the agent has a probe it can dwell in. Canvases so the agent has a workspace in which to accumulate practice. Flows so what the practice discovered can be crystallized and passed forward. None of these are instructional. All of them are — and I mean this in the precise sense Polanyi would have recognized — *personal*.

---

## Coda: what we learned from a chemist-turned-philosopher

Polanyi was a working scientist who turned to philosophy because he had watched his own discipline misrepresent itself. Physicists and chemists *said* they proceeded by hypothesis and falsification, by propositional rigor, by the application of explicit rules. But Polanyi, who had spent his life in laboratories, knew this was not what actually happened. Discoveries were made by feel. Experiments worked because someone had a hand for them. Frameworks were dwelt in, not just used. The official epistemology of science had almost nothing to say about how science was actually done.

Six decades later, we are building something called artificial intelligence and making exactly the same mistake in reverse. We are assuming that if we just write more explicit instructions, more detailed skill files, more carefully designed prompts, the agents will inherit the tacit knowledge of the practitioners whose work they are trying to replace. They will not. The structure of tacit knowledge forbids it.

What Polanyi gives us is both a diagnosis and a direction. The diagnosis: instruction is a weak channel for skill, and most of what separates an expert practitioner from a novice is exactly the kind of knowledge instruction cannot carry. The direction: if you want to build systems that accumulate real competence, you have to build them the way humans actually accumulate competence — through tools that become extensions, through practice that gets reflected on, through crystallizations of what worked that can be transferred without being fully articulated.

We did not know any of this when we started. We were trying to fix a token economics problem and kept finding ourselves building primitives that, in retrospect, corresponded exactly to what a mid-century Hungarian chemist had already described. That is not an accident. Polanyi was working out the structure of knowing, and the structure of knowing turns out to be the same whether the knower is a blind man with a probe, a swimmer in a pool, a surgical resident at Johns Hopkins, or a software agent inside a production system.

We can know more than we can tell. The agent can too — if we build it a workspace where knowing can accumulate, and a way to crystallize what it learns so the next agent doesn't have to start over. That is what skills mean in Rote. Not a document. Not an instruction. A life of practice, made transferable.

---

*Primary sources: Michael Polanyi,* Personal Knowledge: Towards a Post-Critical Philosophy *(University of Chicago Press, 1958), especially Chapter 4 sections 20–25 on tools, frameworks, and the structure of awareness. Michael Polanyi,* The Tacit Dimension *(University of Chicago Press, 1966), especially the first lecture, "Tacit Knowing." Secondary: the Polanyi Society's archive, Nonaka and Takeuchi's* The Knowledge-Creating Company *(1995) on the Matsushita bread-maker episode, and the medical education literature on William Halsted's 1890 surgical residency at Johns Hopkins.*