持而盈之，不如其已。
揣而銳之，不可長保。

A bowl overfilled will spill.
A blade over-sharpened grows dull.

— Lao Tzu, Dao De Jing, Chapter 9

The blueprint is complete. The house already stands within your mind, its pillars named, its beams felt. Yet a house does not raise itself. It needs hands.

What kind of hands? Not new ones.

The craftsman who builds well with language models is the same one who has built well before, in other materials, with other tools, across other seasons of life.

The shift that already happened

There is a moment, quiet and easy to miss, when the nature of your work changes.

For years, I wrote code and shaped systems across languages and lands. The essence remained: translate insight into form. Think clearly, build carefully, iterate and adjust when needed.

Language models did not end this work. They moved it.

Where I once wrote functions, I write intention. Where I once traced logic, I clarify meaning. The rigor never left. The instinct that whispers something is not right never left. All those years of learning how systems break, how clarity fades, how complexity hides—none of it left.

It simply turned.

Many see this craft as something new to learn. It is not. It is where all you already know meets a different kind of clay. The engineer who sensed where code would bend can now sense where a prompt may drift. The architect who felt over-design can now feel an overfilled thought.

You have not started over. You have arrived.

The whole house, held at once

A carpenter does not think about the chisel. She thinks about the piece.

When she planes a board, she feels the grain, watches the edge, senses the humidity in the wood, remembers the joint it must meet. She does not attend to one thing at a time. She holds the whole.

This is what the House of Language Models asks. The foundation, the four pillars, the three trusses, the roof, they are not a checklist. They are layers of awareness, held together.

You shape a prompt while considering how it will be evaluated. You design memory while feeling whether the infrastructure can carry it. You set alignment boundaries while asking if you can observe what happens within them. You choose what data to trust while sensing how the orchestration will move through it.

And above all of it, the roof. Security is not a pillar you build alongside the others. It is what covers the entire house. Every prompt is a door that could be opened. Every agent you deploy, deserves at least the same scrutiny as a human user. To build a beautiful house but leave it open to the rain, is to build no shelter at all. It does not matter how fine the rooms are. The rain will still find them.

This is not mere complexity. It is the natural depth of any mature discipline. A pilot does not fly by one instrument. A conductor does not lead by following one instrument. They hold the whole, and from that calm, precision grows.

The overfilled bowl spills. The over-sharpened blade dulls. To perfect one room while ignoring the roof is to build for display, not for life.

This craft asks for balance. Not mastery in one stroke, but presence in all.

Taste was earned before this

There is a word for knowing why something works, not just that it works. Call it taste.

Taste in language model engineering is real, and it was not born here.

It was built over years of reading a codebase and knowing, before running a single test, where the fault would live. It was shaped by hundreds of code reviews where you learned to see not just what was written, but what was missing. It deepened through debugging distributed systems, where the error was never where you first looked, and the repair never your first guess.

This accumulated judgment, this sense of proportion and balance, is what the discipline actually needs. Language model engineering does not ask for beginners. It asks for people who have already built the instinct and are willing to point it in a new direction.

That is why an experienced engineer, new to this art, may build wiser in weeks than one who has studied patterns for months without depth. Patterns can be learned. Judgment is earned, through seasons, through teams, through stumbles that taught more than success ever could.

The blade does not need more edge. It needs a hand that knows where to cut.

Tools fade, judgment stays

Each year brings new frameworks, new models, new best practices replacing the old.

This is the surface.

Beneath it, the discipline remains. The questions remain. Is this the right data? Is this context necessary or is it noise? Will this evaluation actually catch the failure I care about? Can I observe what matters, or am I measuring what is easy? And the question that outlasts all others: who can reach what, and should they be able to?

A hidden instruction buried in a document that an agent faithfully executes. An access token granted to a workflow because it was convenient, not because it was warranted. These are not exotic attacks. They are consequences of forgetting that the roof must be checked as often as the rooms beneath it. The craftsman who has spent years hardening systems, reviewing access controls, thinking about what an adversary sees when they look at an interface, carries that instinct into this new material without being told.

The craftsman who understands why an output shifted, not just that it shifted, will stand through every change of tools. The one who only memorized syntax will not.

This is why we invest in people, not platforms. For building depth, not chasing trends. In giving makers time to know the whole mountain, not just the path of the season.

Lao Tzu’s wisdom is gentle and clear: A bowl filled past its brim holds less. A blade honed past its nature cuts nothing. There is a point where more learning, more tools, more chasing begins to take away.

The craftsman knows where that point is. Not because she was told, but because she has been there, and learned to stop.

Looking up from the work

For a time, you work alone. Refining the touch, learning to hold the whole house within, growing the quiet confidence that comes from knowing the craft, not the chatter.

But there comes a moment, unexpected, often small, you recognize someone.

A conversation at a gathering. A repository you find by chance. A few lines in a thread that reveal not just skill, but the same inner compass, the same restlessness to build something true, the same quiet refusal to see this as a passing wave.

Someone shaped by a different path, a different tongue, a different sky,
yet carrying the same weight of seasons, turned toward the same new clay.

You recognize the craftsman's mindset before you recognize the person.

That recognition, quiet and certain,
is where the next chapter begins.

This is the fourth article of The Way of The Language Model

This chapter offers that framework, a blueprint laid gently on the table before the first stone is placed.

The metaphor

Every house begins with a foundation. From it rise pillars, strong and quiet supports that bear the weight of the roof and all that lies under it. Between them stretch the trusses, beams that bridge pillar to pillar, sharing the load, keeping the whole from falling. Remove one pillar and the roof begins to sink. Remove one truss and the pillars stand alone, unable to lean on one another.

The House of Language Model is built this way. It is not a list to check, but a living structure, each part leaning into the next, each breath held in balance.

The data foundation

The foundation is data, it's the base plate upon which the house sits.

Every language model, however wise it seems, rests upon the data it learned from and the data it meets in each new moment. Training data shapes what it knows. Inference data shapes what it can do. If the foundation is weak, unsteady, unclear, untrue, then the house trembles upon sand.

This work is quiet. It rarely draws applause from stages. Yet the builder who neglects it will spend years mending cracks that never truly heal.

The four pillars

The pillars are the steady disciplines that hold the house upright. They are the quiet work that surrounds and sustains the model’s spirit.

Orchestration is harmony. When an agent must call a tool, look into memory, walk through reason, and then speak, something must conduct the flow. What comes first? What follows? Orchestration brings order. Without it, cleverness wanders without direction.

Evaluation is reflection. How do you know the answer is true? And not only the answer, was the path it took also clear and right? A single question may be simply right or wrong. But an agent that browses, asks, wonders, and finally offers a gift, this calls for a deeper kind of seeing. Without evaluation, growth is but a guess.

Infrastructure is the earth below the floor. Where does the house stand when ten thousand guests arrive at once? How does it balance weight across lands, shelter uncertain code, welcome each one by name? Infrastructure is invisible when it blooms. When it breaks, nothing else remains.

Observability is sight. When an agent falters, can you see why? Can you trace its steps, find where it turned, and understand its hesitation? A system you cannot see is a system you cannot trust, for you cannot learn from what stays hidden.

These four pillars do not give answers. They prepare the ground where answers may grow, steady, lasting, and true.

The three trusses

If pillars hold the house up, trusses weave them together. They are the gentle arts, the daily practices of those who tend to language.

Invocation is the art of prompt engineering. A language model reads only forward, rarely backwards, like a stream that never turns. Thus the order, the tone, the clarity of your words are not mere preference. They shape the world the model sees. How you call determines all that follows.

Memory is the discipline of context engineering. A model holds no past of its own. It does not recall your last conversation, your quiet rules, your intentions, unless you place that knowing within its reach. Memory is choosing what must be remembered, and offering it in a form the model can hold. Too much, and the signal is lost in noise. Too little, and the model walks in twilight.

Honor is the practice of alignment engineering. A model that speaks smoothly but ignores your boundaries is not only unhelpful, it may bring harm. Honor is the discipline of gentle guardrails, ensuring the model moves within the garden you have tended, respecting safety, ethics, and your deepest intent. It is a promise that the system does not only what it can, but what it should.

The trusses bind the pillars together.
Invocation without evaluation is a song without an ear to hear it.
Memory without infrastructure cannot travel far.
Honor without observability is a promise made in the dark.
The house stands because these arts lean into one another, each breath shared, each weight balanced.

The roof

A house without a roof is open to the sky. Beautiful, perhaps, in summer. Ruinous when the storm arrives. Security is the roof of the House of Language Models. It does not draw the eye the way a pillar does, nor carry the intimacy of a truss. Yet it covers everything. Every pillar, every truss, every room beneath it shelters under its care.

A language model system faces rain from many directions. Prompts may be twisted by those who wish to reach beyond the garden walls. Data may be poisoned at the source, quietly, long before the builder notices. Outputs may leak what was meant to stay hidden. The roof must hold against all of these, not with a single beam, but as a continuous canopy woven across the whole structure.

Security is not a gate you place at the door. It is a way of thinking that touches every choice. How is the model called? What may it remember, and what must it forget? Who may speak to it, and what may it speak of? These are not afterthoughts. They are questions the builder asks with every stone she lays.

A roof built last, hastily nailed over a finished house, will always have gaps. A roof imagined from the first sketch, drawn into the blueprint beside the pillars and trusses, becomes part of the silence that makes the house feel safe.

And so the roof does not appear as its own chapter in the rooms ahead. It appears in every chapter, a quiet presence overhead, reminding the builder that shelter is not optional.

The House of Language Models

It is tempting to love one room and forget the rest. A builder skilled in prompts but blind to evaluation may craft wonders that shine in a demo and fade in the world. A team that tends only to infrastructure but forgets alignment may scale a house that should have remained a small, kind shelter.

The House of Language Models is one complete practice. Each chapter ahead will step into one room, study its windows, touch its walls, and see what a thoughtful hand may build there. Yet always, the house is whole.

And so the next chapter asks a different kind of question. Now that we hold a blueprint, what kind of person builds from it? What grows in the heart of the craftsman?

This is the third article of The Way of The Language Model Engineer

三十辐共一毂，当其无，有车之用。
埏埴以为器，当其无，有器之用。
凿户牖以为室，当其无，有室之用。
故有之以为利，无之以为用。

Thirty spokes share one hub. Where the wheel is not, that is where the cart is useful. Clay is shaped into a vessel. Where the vessel is not, that is where the vessel is useful. Doors and windows are cut for a room. Where the room is not, that is where the room is useful.

Therefore, what is present creates benefit. What is absent creates usefulness.

– Lao Tzu, Dao De Jing, Chapter 11

When you begin working with language models, there is an instinct to fill the space.

You write longer prompts. You add more context. You craft detailed specifications for every corner of a repository, every rule, every preference. It feels natural: more information, you assume, will bring better results.

For a moment, the potter thinks the same. More clay. Thicker walls. A heavier, more solid vessel. Yet in the end, it cannot hold water to drink, rice to eat, or tea to share. It holds nothing at all.

Twenty-five centuries ago, Lao Tzu understood. A vessel's usefulness lies not in the clay, but in the emptiness the clay creates. A wheel turns because of the hollow at its hub. A room shelters because of the space within its walls. Structure exists to give shape to the void.

This is the heart of language model engineering. Not the tools, not the frameworks, not the specifications themselves, but the space they leave for intelligence to arise.

The Stuffed Context

Early in 2026, researchers at ETH Zurich published a quiet revelation. They studied a common practice: giving coding agents repository-level context files, documents like AGENTS.md, describing a project's conventions, architecture, and preferences. Such files had become standard. Teams spent hours perfecting them.

The finding was gentle, yet clear. Across models and agents, these files tended to lower success rates, compared to offering no repository context at all. Meanwhile, inference costs rose by over twenty percent.

More clay. Thicker walls. A vessel that could hold nothing.

In the traces of behavior, something appeared. The context files did shift how agents acted: they explored more, opened more files, ran more tests. They were busy. They were thorough. And they were less effective. Extra, unnecessary demands made tasks harder. The lesson was soft but certain: context should carry only what is essential. ¹

Another study, from Google Research, quietly echoes this. For models that do not reason step-by-step, simply repeating the prompt twice improved performance across the board, outperforming in 47 out of 70 benchmarks, with no losses. ² Why? In causal language models, tokens can only attend to what came before them. When context comes before the question, the model may lose sight of what is being asked. Bury the task, and clarity fades.

This is the stuffed vessel, seen from two sides. Too much context drowns the purpose. Even well-placed context, if the task waits at the end of a long trail, grows distant. The vessel serves not because it is filled, but because it stays open.

This finding did not surprise me. I had already experienced it.

From Overhead to Emptiness

I started using GitHub SpecKit, a thorough system for repository context, promising to give AI agents full understanding of a codebase. The specs were detailed: dependencies, architecture, patterns, tests. By any ordinary measure, they were well made.

And yet, my agents began to drift.

Not suddenly. Not obviously. But over time, a quiet shift: specification drift. Token costs crept up. Replies grew longer, yet said less. The agents were processing more, understanding less.

I recognized this. I had seen it before in software teams. The issue was seldom too many details; it was details set without care, written in haste, without knowing what truly mattered. Thick documents were shipped not from deep thought, but from lack of time to find the core. Noise wore the mask of thoroughness.

With language models, this dynamic grew stronger. Give a model too much context, and it does not align more closely with your aim, it drifts away from it.

So I shaped something different. The Fullstack Requirement Specification³, an open format resting on one idea: not too detailed, not too broad. Only the essential form of intent.

An FRS begins with a little structured metadata: who the user is, where they begin, what starts the action, what they hope to find. Then a numbered flow: the happy path, four or five steps, no more. Alternate paths rest gently beneath, marked with dashes. Optional technical constraints follow, only if they exist. A validation section closes the circle.

That is all. The full spec for an authentication flow fits on one screen. A human reads it in a minute. An AI parses it without strain.

What FRS leaves out matters more than what it holds. It does not describe implementation. It does not impose architecture. It does not explain the philosophy behind choices. It does not offer system-wide context unless that context is a direct need. It cuts doors and windows. The room inside, where the real work happens, is left purposefully empty.

Yet the validation section deserves special care. FRS ends with test cases that form a quiet contract: happy path checks, edges, invariants, logical promises. The AI receives the spec, builds the implementation, and validates its own work by running the tests. If they pass, the work is complete. No human handoff for code review, the tests are the handoff. The specification holds the intent. The validation shows the intent was met. The circle closes.

The change was immediate. Agent performance improved not because I gave better instructions, but because I gave fewer. The light core offered just enough shape for intelligence to fill the space. From that stillness, I could adjust, add a constraint here, refine a boundary there, without being weighed down by everything at once.

Begin with a clear core. Build step by step. Let the emptiness do its work.

The Principle

This is not mere minimalism. The Dao De Jing does not simply argue for less. It observes, with grace, the bond between form and function.

The thirty spokes are needed. Without them, there is no wheel. Yet the spokes exist to make the hub, the empty center around which all turns. Remove the spokes, you have nothing. Fill the center, you also have nothing. The art lies in knowing what to shape and what to leave open.

In language model engineering, this moves through every layer.

A prompt is a vessel. Its value comes not from how much you put in, but from how clearly you mark what is needed. The finest prompts gently bound a space where the model's capacity can breathe. The weakest are walls without openings, so full that intelligence has no room to move.

A context window is a room. Its worth comes not from how much you place inside, but from how well you choose what belongs there. Each needless token is a brick where a window should be. The right context, minimal, relevant, well-ordered, opens the way for understanding.

An alignment constraint is a doorframe. It does not fill the passage. It outlines the boundary that lets one pass through. Without it, there is no door, only a wall, or an open field. The constraint creates the useful emptiness of safe, guided action.

This is what the ETH Zurich researchers measured, though they may not have named it so. The context files were clay. The agents needed vessels.

What Comes Next

If this principle holds, and the wisdom of ages, and the quiet data of today, say it does, then language model engineering is, at heart, a gentle art of taking away. We do not build by adding. We build by shaping.

Yet shaping asks for a form. Emptiness without walls is not a room, it is nothing. The vessel needs its clay. The wheel needs its spokes. The room needs its frame.

In the next chapter, we shall meet that frame. The House of Language Models gives emptiness its shape: three pillars that mark where language model engineering truly lives. Not as a structure to fill, but as walls that make a room worth dwelling in.

The craftsman does not start with ornament. She begins by asking: what must this space hold? And then she builds only what the holding needs.

This is the second article of The Way of the Language Model Engineer.

¹ Gloaguen, T., Mündler, N., Müller, M., Raychev, V., & Vechev, M. (2026). Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? arXiv:2602.11988.

² Leviathan, Y., Kalman, M., & Matias, Y. (2025). Prompt Repetition Improves Non-Reasoning LLMs. arXiv:2512.14982.

³ Lundquist, Decerno AB. https://github.com/openspecs/frs

千里之行，始於足下

A journey of a thousand miles begins beneath the foot.

— Lao Tzu, Dao De Jing, Chapter 64

Before you walk, the ground is already there.

In 2017, Google researchers introduced the transformer architecture in the paper Attention Is All You Need. A quiet seed was planted. Most of the world took little notice.

Years passed in steady, unseen labor—refining, scaling, testing. The tree grew in silence while the world looked elsewhere.

Then came late 2022 and ChatGPT. Overnight, a research curiosity became a global conversation. Language models moved from obscure to omnipresent. Anthropic advanced safety, DeepSeek proved open models could rival closed ones. In two years, wonder turned to utility.

This wave is different.

The internet connected people through technology. Language models do something deeper: they connect technology to language itself. For the first time, the medium of human thought and the medium of computation are merging. This is not just another tool, it reshapes the relationship between how we think and how machines work.

Having lived and worked across continents, Barcelona, Copenhagen, Berlin, Paris and Buenos Aires. I’ve felt how language shapes thought, culture shapes language, and technology reshapes both. When language models emerged, I saw more than a trend. I saw the convergence of language, culture, and technology into one discipline.

So I dedicated at least a full year understanding it, deeply and wholly.

After 17 years in tech, including seven at Volvo Cars working across languages and cultures, I thought I knew what transformation felt like. But this shift changed the nature of the work itself. My craft has moved from writing code to writing precise specifications. The engineer’s role is evolving from implementation to articulation and intent, defining what a system should do so clearly that a language model can execute it well.

The journey begins beneath the foot, not ahead in the distance, but where you already stand. Every engineer brings years of experience, intuition, and hard-earned wisdom. Language model engineering does not erase that foundation; it builds upon it.

That is dedication. Not mere enthusiasm or curiosity, but a professional commitment to understand how things work, not just how to use them. To study the architecture, the alignment, the orchestration. To build, fail, and refine. To honor the craft with the seriousness it deserves.

And yet, dedication extends beyond ourselves.

The next generation will carry the entire internet in their pocket, not as distraction, but as a capability. Language models will transform healthcare, education, and infrastructure in ways we are only beginning to imagine. The areas most vital to humanity are where this technology can serve best.

We dedicate ourselves now so that what we build is worthy of their trust.

This is where the journey begins. Beneath the foot. In stillness.

This is the first article of The Way of the Language Model Engineer.