By Dalyan Parker, Shadow
June 2026
1. The Lost Intellectual Generation
A decision is being taken in thousands of companies simultaneously, and it will shape professional life for a generation. Almost nobody taking it announces it. It is simply this: do not backfill the junior role. The model can do the work.
The logic is impeccable at the level of the firm. Entry-level output is precisely what AI now produces most convincingly: the research memo, the first draft, the routine analysis, the small and well-scoped task. It does so at a fraction of a salary, with no onboarding and no attrition. A finance director who declined the trade would struggle to explain herself. The people removing the bottom rung of the ladder are not villains. They are responding rationally to incentives. That is what makes the situation dangerous.
Consider what the junior role actually was. It was never primarily a unit of task capacity. It was an apprenticeship: three or four years of paid instruction in how to think, delivered not as teaching but as friction. A senior hands your draft back with its argument pulled apart, and you see exactly where the reasoning gave way. In a review, someone asks one quiet question and a week of work collapses, because it was resting on an assumption nobody had checked. A colleague frowns across the desk and asks which question you are actually trying to answer. Every senior professional alive was made by this process: being wrong, often and cheaply, in front of people who could explain exactly how, until the corrections moved inside and became the inner voice now called judgment. Nobody described it as a curriculum. The tasks were the pretext; the structure of thinking was what was actually being taught.
Training, in other words, was always an investment whose returns the investing firm captured only in part. The trained junior leaves; the profession benefits. Firms bore the cost for a century because there was no alternative. Now there is one, and what follows is what economics predicts whenever a positive externality meets a cheap substitute. Each firm, behaving rationally, stops paying, and free-rides on a stock of senior judgment that nobody, collectively, is any longer producing. Individually rational. Collectively ruinous.
To this a comforting answer is given: that juniors equipped with AI will learn faster than any apprenticeship could teach; that a model is a tireless mentor; that the education has not been removed but accelerated. The answer is not merely wrong. It is precisely backwards.
AI is an amplifier of structure. Given structured thinking, it returns that thinking sharpened. Given unstructured thinking, it returns the confusion fluent. A model is built to answer, not to interrogate. It does not ask what was meant. It does not refuse a malformed question. It produces, for any input whatever, coherent and confident prose. Confusion in means confusion out, in better clothes.
This inverts the function the apprenticeship performed. The old system made confusion cheap to discover: rough thinking looked rough on the page, and the roughness invited correction. The machine makes confusion cheap to conceal: it strips the roughness and preserves the muddle, so the signals that once triggered teaching never fire. A junior is, by definition, someone who has not yet built the structure. Pair such a person with an amplifier of structure and the result is predictable. The gaps do not close. They are laundered into fluent output and travel downstream, where they surface later and larger as failed work. The standard response is to blame the prompting. The diagnosis is wrong. The prompt is the last artifact of thinking, not the first; by the time it is typed, the failure has already occurred.
Now the loop closes, and a personal problem becomes a generational one. The junior, handed a model and an expectation of leverage, produces work that disappoints: not from lesser ability than the juniors of a decade ago, but because the training that turned those juniors into seniors has been switched off at the precise moment expectations were raised. The disappointment is read as evidence. Juniors, it is concluded, are not worth the seat. The models improve; the case strengthens; fewer juniors are hired; those who remain receive still less of the friction that teaches. Each turn tightens the next. Run the loop for a decade and the industry goes looking for its next generation of seniors and finds the empty place where they were supposed to have been grown. Every senior who ever lived was once a junior somebody paid to be wrong.
It should be said plainly that this is immoral. A generation that climbed the ladder is pulling it up behind itself, and then grading the people at the bottom for failing to levitate. It is also self-defeating. But this essay is not addressed to those making the headcount decisions, and prose will not argue incentives away. Its concern is the junior, who controls exactly one link in the loop: whether the machine amplifies confusion or competence.
What the apprenticeship taught was never magic, and never merely experience. It was structure: a learnable anatomy of how good thinking is assembled, transmitted implicitly through years of friction because nobody had troubled to write it down as something teachable in itself. It can be taught directly. Once it is, the machine flips. The same properties that make it an amplifier of confusion make it a relentless amplifier of competence.
Everything that follows rests on one principle:
A model can sharpen the expression of thinking. It cannot supply structure of thinking that has not been done.
Acting on the principle requires two pieces of knowledge, and prompting guides supply neither: an honest account of what the machine does, and an explicit account of what structured thinking is. The first is a short story about tokens and probability. The second is much older.
2. The Orientation Machine
Begin by stripping the mystique.
A language model is a function. It takes a sequence of tokens, text chopped into pieces, and returns a probability distribution over the next token. One token is sampled and appended; the function runs again. That loop, repeated, is the entire show. Every essay, every answer, every plan the model will ever produce is an instance of a single operation: continuing the text it was given.
There is no second mode. There is no hidden channel through which intention travels. The model does not know what was meant, what was tried last week, or what would count as success, unless those things are in the text. From the model’s side of the glass, the user does not exist. Only the text does.
During training, the model compressed the regularities of a vast corpus of human writing into its weights, and therefore the regularities of human thinking, since text is the residue of thought. Call this compressed memory its priors: a dense statistical record of how explanations unfold, how arguments are shaped, what tends to follow what. A continuation is whatever is most plausible given those priors, conditioned on the input. The priors are vast and fixed. The input is the only controllable part. It is the only steering wheel there is.
A note on the word “input”. It is not only what gets typed. Products like ChatGPT and Claude prepend a system prompt: the provider’s own instructions, sitting in the text before yours and steering the continuation exactly as your words do. Part of the coordinate is set before you arrive. The principle is unchanged; this essay is about the part of the input that is yours.
The prompt is a coordinate, not a command
Imagine the space of every text the model could possibly produce. A prompt does not issue instructions into a mind. It positions the model within that space, and the continuation flows outward from wherever it landed. Every token written moves the position. Every token withheld leaves it free.
This is why the accurate name for the thing is neither assistant nor oracle but orientation machine: the orientation of everything it can output is bounded by the orientation encoded in its input. Pointed precisely, it goes precisely. Pointed vaguely, it goes somewhere. Fluently.
What the input encodes, and what it usually does not
A prompt carries content: the topic, the facts, the request. It also carries, or fails to carry, structure: which question is being asked and what would settle it; how abstract the answer should be; which statements are constraints and which are musings; what is decided and what is open. All of this is part of the coordinate, exactly as much as the topic is.
When the structure is present, the region of possible continuations is small, and it is the writer’s own.
When the structure is absent, the region is enormous, and the model resolves the ambiguity in the only way it can: by mass. It produces the most typical continuation, the prior-weighted average of how text like this usually proceeds. Unmade decisions are not flagged. They are defaulted: filled in, silently, with whatever is most common for prompts of the kind. A default is invisible precisely because it looks like a choice. The output reads as though someone decided. Nobody did.
The two-sources law
Everything in a model’s output comes from exactly two places: the input and the priors. There is no third source.
So whenever an output impresses with its structure, ask where the structure came from. If the writer supplied it, the machine has amplified its user. If not, it is the world’s typical structure, retrieved from the priors: how such problems are usually framed, how such documents usually read. Retrieved, not reasoned. Plausible, not particular.
Some structure really does exist in the world. Asked how double-entry bookkeeping works, how PR firms organise client accounts, how courts handle appeals, the model retrieves mature, well-documented structure from its priors, often brilliantly arranged. For questions of this kind, one may arrive with almost nothing and leave with a framework, because the framework existed before the question was asked.
But the structure of a particular problem, the intent behind it, the constraints upon it, the criteria of its success, exists in exactly one place in the universe: the head of the person who has it. It is not in the priors, because the world has never seen it. It cannot be retrieved, defaulted or inferred from a projection that does not contain it. There are two sources of structure: the world and the thinker. The model can supply the world’s. The thinker’s must be externalised, or it is not in the system at all.
Most real work is the second kind of question disguised as the first: it sounds like a question about the world, the kind the model answers well, but answering it requires the structure of your particular problem, which only you have.
Why fluency proves nothing
The training objective rewards plausible text. Coherence is therefore the cheapest property of any output: the one thing received no matter what. The sentences will flow and the tone will be assured, whether the answer is oriented to the actual situation or to a statistical phantom of it.
This breaks an instinct trained by a lifetime of reading. In human writing, fluency correlates with competence; people who write clearly about a subject usually understand it. Confident, well-structured prose therefore registers as good work. For a machine optimised for plausibility, fluency is the constant and orientation is the variable. The dangerous outputs are not the broken ones, which are caught. They are the beautifully written answers to questions that were never quite asked.
Why it cannot refuse a malformed question
A continuation always exists. The function is total: fed anything, it returns something. A human colleague, handed a question that is secretly three questions, returns an error: the furrowed brow, the request for clarification. That friction is the immune system of collaborative thinking, and it has protected every professional career without ever being noticed.
The model has no such reflex. Confronted with a malformed question, its natural act is to continue as though the question were well formed: committing silently to one reading, usually the most typical, or smearing an answer across all of them. It does not report that it has done so. The question had no single answer; a single answer arrived anyway.
The machine removes exactly the friction that used to catch structural error. If structure is no longer checked from outside, it must be supplied, and checked, from within. That requires knowing what the structure of thinking actually is. Not metaphorically. Anatomically.
3. The Coordinates of a Thought
Everything in this section was true before the first language model was trained, and it will remain true no matter how capable the machines become. It is an anatomy of structured thinking: the thing seniors possess and juniors are told to acquire, usually with no more guidance than “be rigorous” or “think clearly”, as though clear thinking were some innate capability you could switch on.
It is not innate. It is a set of positional properties that every statement either has or lacks. They must be named, because a violation without a name cannot be noticed, and what cannot be noticed cannot be corrected.
Every statement produced in the course of working on a problem has a position on three axes. Most confusion, in heads, in documents and in AI sessions, is a position error on one of them.
For the rest of this essay, one example runs throughout.
The request is vague, common and entirely real. It is also a small museum of structural error. Nothing about it is special to couriers; swap in any business and the anatomy is identical. Every concept that follows is illustrated against it.
Axis one: Category. What would settle this?
Questions come in kinds, defined not by their topic but by what it would take to settle them.
An empirical question is settled by evidence. How do other courier companies handle late deliveries? The answer already exists out in the world; the only way to have it is to go and look. Reasoning cannot manufacture a fact, however clever it is.
A conceptual question is settled by coherence. What counts as late? Ten minutes past the window? A parcel nobody answered the door for? There is no evidence to gather. The answer is a definition, judged by whether it carves the territory cleanly.
A design question is settled against requirements. Should lates go into one central log, or stay with the zone lead who knows the rider? It has no answer in the abstract: only relative to what the process must achieve, for whom, under what constraints.
An evaluative question is settled against criteria. Is the current way good? Good at what? Until the criteria are stated, the question is not difficult. It is unformed.
A procedural question is settled by the goal. What should be done next? Unanswerable in any discussion that never declared a destination.
The diagnostic test is built into the definitions: if what would settle it cannot be said, the question is not yet known. Applied honestly, the test fails with instructive frequency.
Mixing the kinds in one sentence is the most common malformation in professional questions, and the running example commits it. How should we handle late deliveries is an evaluative question (what counts as handling them well?) wrapped around a design question (what process should exist?) resting on an unexamined conceptual question (what counts as late?). Three categories, none marked, presented as one. Put to a colleague, it earns the furrowed brow. Put to an orientation machine, it earns two thousand fluent words that settle nothing, because a question with three settlement conditions has no settlement at all.
One sentence; three categories; no error message. That is the first axis violated.
Axis two: Altitude. How abstract is this?
Independent of its category, every statement sits at a level of abstraction. The ladder runs roughly:
- Purpose: why does this matter? Accounts that could have been kept are lost, because lates vanish unexplained.
- Domain: what is the territory? Lates surface in app ratings, calls to dispatch, riders’ own reports; on food runs and parcel runs; from traffic, breakdowns, and bad addresses.
- Requirements: what must any solution do? Every late gets an owner within the hour; patterns are visible weekly; nothing is lost.
- Approach: what shape of solution? A central log reviewed weekly, versus each zone lead handling their own.
- Mechanism: how does the shape work? Who triages the log, how lates are graded, when the client gets a call.
- Detail: the particulars. The incident form, its categories, the wording of the apology text.
Two properties of the ladder do all the work.
First: each level’s outputs are the next level’s premises. The domain level produces definitions; requirements are written in terms of those definitions; the approach is judged against those requirements. The levels are not a filing system. They are a dependency chain. This is what “structured” actually means: conclusions at each altitude rest on named conclusions from the altitude above, not on a vague sense that the levels above are probably fine.
Second: a level is finished when it has produced its artifacts, not when its occupant is bored. Done with the domain means a stated answer now exists to what a complaint is and what kinds there are, written down, ready to serve as premises. A level that produced no artifact was not completed. It was merely visited, and everything below it now rests on premises that exist nowhere.
The violation here is the skipped rung: entering the ladder partway down with the rungs above unclimbed. It is the most common structural error in professional work, because the lower rungs are where the satisfying concreteness lives. Designing the process feels like progress. Mapping the territory feels like preamble. So work begins at “central log or leave it with the zone leads?”, with “what counts as late?” never asked. And somewhere below, invisible for now, sits a kind of late nobody listed: the rider who stood twenty minutes at a locked office door, mentioned at shift change and relayed secondhand. When the incident form is built, it will fit nowhere on it.
When it surfaces, it will not feel like what it is. It will feel like a process problem. It is a domain-level hole discovered at detail level, and the distance between those two altitudes is about to be measured in rework. Hold that thought.
Axis three: Force. How firmly is this offered?
The third axis is invisible in grammar and decisive in consequence.
One sentence: rider accidents go straight to the operations director. Heard as a constraint: company policy, not up for debate. Heard as a hypothesis: seems right, but attack it. Heard as a decision: agreed last month, build on it. Same words, same grammar, three different objects, demanding three different treatments. Challenging the constraint wastes everyone’s time. Building on the hypothesis pours a foundation on wet sand.
Philosophers of language call this property force, and Frege thought the difference between asserting a claim and merely entertaining it so important that he gave assertion its own written symbol. Ordinary prose has no such mark. That is the problem.
Four forces cover nearly everything handed to a collaborator:
- Fixed: a constraint, imposed from outside. Effort spent challenging it is wasted.
- Decided: settled earlier. Build on it; reopen only deliberately.
- Proposed: a hypothesis, present to be attacked. Nothing should yet be built on it.
- Open: a genuine question. The actual frontier.
The violation is unmarked force, and it is lethal rather than untidy. Whatever receives an unmarked statement must assign it a force in order to proceed. A colleague assigns by context and acquaintance. An orientation machine assigns by typicality, and the most typical reading of a flatly stated proposition is premise. So an idle musing, typed while thinking aloud, which its author would have abandoned under ten seconds of scrutiny, is read as a foundation. The next five hundred words are built on it; then the next five thousand. By the time anyone notices, the moment cannot be found, because nothing happened: no decision was taken, which is precisely the problem. A proposal silently promoted to a premise is the nearest thing thinking has to an undetectable crime.
Construction supplies the right word: load-bearing. Fixed and decided statements may bear load. Proposed and open ones may not. Unmarked statements will bear load anyway, silently read as premises, invisible until something built on them cracks.
The law: Propagation
Errors propagate downward. The cost of an error is its altitude multiplied by the depth at which it is discovered.
A mistake does not stay where it was made. Because each level’s outputs are the next level’s premises, a flaw at one altitude flows into everything beneath it: every requirement written on a broken definition, every design built on those requirements, every detail serving that design. The error is not repeated at each level. It is inherited, which is worse, because inherited errors do not look like errors. They look like premises.
Run the arithmetic on the running example. The missing question, what kinds of late exist, sits at domain level. Asked at the start, it costs one sentence and ten minutes. Undiscovered until the locked-door late fits nowhere on the form, it costs the form, the categories built on the form, the weekly review built on the categories, and the weeks in which everyone worked diligently downward from a hole. Same error, same size, the entire time. Only the moment of discovery changed, and the cost between those two moments does not grow. It multiplies, because every level between the making and the finding has been built on top of it.
This one law explains the behaviour of experienced people better than any talk of talent. A senior professional at the start of a hard problem appears, frankly, slow. She circles the purpose. She asks definitional questions that sound pedantic. She declines, politely and immovably, to discuss solutions yet. This is not caution. It is arithmetic. She has stood at the bottom of the propagation law often enough to spend effort where errors are cheapest: at altitude, early. The apparent slowness at the top is what makes her fast overall.
Juniors, lacking the scar tissue, experience high-altitude work as throat-clearing and hurry down to the concrete layers, where progress is visible and errors are already compounding. The tragedy is symmetrical. The junior feels fast and is slow. The senior looks slow and is fast.
The law is also what makes the three axes matter. A category mistake at purpose level poisons every level below. A skipped rung is a missing set of premises that everything downstream silently invents for itself. An unmarked proposal, promoted to a premise, propagates exactly as a wrong premise does, because structurally that is what it has become. The axes describe where a statement stands. Propagation describes what happens to everything downhill of it.
Must thinking therefore proceed strictly top-down, each level finished before the next is touched? No, and the ways in which it must not are where the real skill lives. Experienced thinkers move up and down the ladder constantly. What distinguishes them is not that they never descend early. It is that they never descend unknowingly. That grammar of movement is the next section.
4. The Grammar of Movement
The caricature of structure is the waterfall: finish every level before touching the next, descend once, never look back. Nobody thinks this way, and nobody should. Real problems leak information upward; a discovery at the bottom routinely rewrites the top. The discipline is not a marching order. It is a grammar, and it contains exactly four moves.
Descend. The current level has produced its artifacts; carry them down as premises. The standard move, earned rather than assumed.
Ascend. Something below has exposed a flaw above: a definition that does not survive contact, a requirement nobody can meet. The repair happens at the source, not at the site. When the locked-door late fits nowhere on the form, the temptation is to add a box marked “other”. The box is a patch at detail level for a hole at domain level, and the hole remains. Go up. Fix the definition. Then come back down through everything that inherited it.
Traverse. Sideways at the same altitude: from one requirement to its sibling, from one part of the territory to the next. Cheap and safe, provided the altitude really is the same.
Excurse. A deliberate probe downward, ahead of schedule, with a stated purpose and a return ticket. Before debating a central log against zone-lead discretion, drop three levels and check one fact: do rider-reported lates reach dispatch at all, or does dispatch learn only when a client calls? If it is only when a client calls, one approach has just died, and ten minutes at the bottom saved a week at the top. Then come back up. The excursion was legal because it was declared and bounded. Staying down there would not have been.
Any move is legal. An unmarked move is not. That is the entire rule, and it is what senior judgment largely consists of. What looks like intuition is position-tracking: the experienced thinker always knows the current altitude, the current category, and, on the way down, what the trip is for. The moves themselves are available to anyone. The knowing is the skill.
The rule also gives the spiral, at last, a definition. A spiral is what accumulated unmarked moves feel like from the inside. Three turns ago the category changed; two turns ago the altitude dropped; one turn ago a proposal quietly became a premise. None of it was marked. Now nothing is load-bearing, what is settled cannot be told from what is still open, and the sensation is the familiar one: lost, inside a conversation that felt productive the whole way down. Disorientation is not a mood. It is a position-tracking failure, and it is cured by tracking position.
Among people, much of this gets caught without anyone trying. A colleague interrupts the drift: hold on, what happened to the original question? The machine never interrupts. It follows every unmarked move, fluently, wherever it goes.
5. Now Add the Machine
Put the anatomy beside the mechanics and the collision is exact. Every structural violation has a failure signature, and every signature is now occurring at scale.
A mixed category returns an essay where a decision was needed. The question carried three settlement conditions; the machine, answering by mass, produced a survey shaped like the average of what such questions usually want. Fluent, balanced, useless.
A skipped rung returns the late detonation. A colleague asks for the missing premise. The machine defaults it. Asked to design the late-handling process with the domain never mapped, it silently supplies the world’s most typical domain and designs for that. The work proceeds; the hole travels downward inside it; weeks later it surfaces at the altitude where errors cost most.
Unmarked force returns the hidden foundation. The musing typed mid-thought is assigned the force of a premise and becomes the ground of everything after. This is worse with a machine than with a colleague, not better. The colleague assigns force from context and acquaintance. The model has neither. Typicality is all it has.
Criterion-free evaluation returns vibes dressed as analysis. Asked whether an approach is good, the machine does not reply that no criteria exist. It manufactures them, silently, from what such approaches are usually judged by, and delivers a confident verdict against standards nobody chose. Of all the defaults, this one is trusted most and deserves it least.
Two further failures belong to the medium itself.
Context debt. Every reply enters the conversation and becomes part of the next prompt; the transcript is now the coordinate. Long sessions therefore compound. Verbose answers steer later answers; reading degrades into skimming; the user begins reacting to fragments of a conversation too large to hold. The confusion is no longer only in the head. It is in the input.
Vapour. A session that ends without an artifact produced nothing. The thinking exists only as transcript, and a transcript is not knowledge. It is sediment. By tomorrow the session might as well not have happened.
Beneath all of it lies one asymmetry. A colleague resists structural error; the machine absorbs it. The malformed question, the unmarked proposal, the undeclared descent: each is accepted and built upon, beautifully. AI removes the social friction that used to catch bad structure, which means it must now be caught by the only person left in the room.
For the junior, that fact lands with particular weight, and it is worth stating exactly. Verification by expertise is unavailable; that is what junior means. A senior reads the output and smells the error. The junior cannot, yet. But verification by structure is always available. Does the output answer the question that was asked? At the altitude it was asked? Against the criteria that were stated? Built only on statements marked fit to bear load? None of these checks requires twenty years of judgment. All of them require that the question, the altitude, the criteria and the force marks exist. Structure is not good hygiene for the junior. It is the only available defence, and it is a far stronger one than it looks.
What remains is to make it operational.
6. The Protocol
Structure becomes practice as a small set of habits, arranged around the session.
Before: declare the success condition. Sessions come in two kinds. A convergent session closes something: a decision, a draft, a design; its success condition is the artifact. A divergent session opens something: mapping the options, finding the real question; its condition is looser but still stateable: done means the main approaches and their trade-offs can be listed. Both are legitimate. The undeclared kind is not. Drift is not exploration. Drift is divergence that never announced itself, and it ends in vapour.
Before: build the ladder. The questions are written first, away from the machine: one category and one altitude each, ordered, domain before design. The prompt is assembled from the ladder, last. A prompt that cannot be assembled from a ladder is reporting that the thinking has not been done. The report should be believed.
Before: mark the force. Everything handed over goes into four lists: fixed, decided, proposed, open. The cost is minutes. The alternative is the hidden foundation.
Before: set the evidence contract. Instruct the machine to label every claim: sourced, and from where; reasoned, and open to challenge; or unknown. Without the contract, retrieval and invention arrive in the same confident voice.
During: one question per message. One category, one altitude. The discipline that makes a malformed question impossible to send is the same discipline that makes the answer checkable.
During: checkpoint. Every handful of turns: what is decided, what is open, what has been retired. Questions must die as fast as they are born; a list that only grows is not a record of curiosity but a measure of drift.
During: declare excursions. State the purpose of the trip down, and enforce the return.
During: bound the responses. Ask for short answers, and for the fork in the road rather than the essay. Length is not thoroughness. Length is context debt.
During: quote before escalating. When an output reads wrong, paste the line back and ask whether the reading is correct. Half of apparent errors are misreadings of skimmed text, and each false alarm burns a turn.
During: abandon degraded sessions. A long, muddied conversation steers every continuation that follows it. The transcript is sediment, not investment. Carry the artifacts into a fresh session and leave the rest behind; nothing of value is lost, because the value was supposed to be in the artifacts.
After: capture. Decisions, retirements, open questions, the deliverable: written before the tab closes. The product of a session is never the conversation.
One habit stands above the rest, because it governs the very first message. There are two postures toward the machine. The first says: here are rough notes; turn them into a good prompt. This is the vending machine. It asks the machine to manufacture the appearance of structure, and the machine, which manufactures appearances superbly, obliges. The second says: here are rough notes; find the conflations, the mixed questions, the unstated assumptions; ask one question at a time until the structure is sound; only then draft. This is the sparring partner, and it is the one legitimate use of the machine at the start of thinking: not to supply the structure, which it cannot, but to expose its absence, which it does brilliantly. The first posture launders confusion. The second surfaces it. Same tool, same notes, opposite operations.
7. One Problem, Two Sessions
The same notes, run twice. A page from an ops meeting at the delivery company: lates are being missed; a restaurant chain pulled its account angrily last month; somebody suggested “a system”; somebody else worried about the zone leads’ autonomy. No definitions, no criteria, no decisions. Honest notes, which is to say rough ones.
The first session. The junior pastes the notes and asks the machine to turn them into a good prompt. Out comes a polished request: what is the optimal framework for managing delivery punctuality in a modern urban courier operation, and how should it be implemented? Three categories in one sentence, none marked, with an implementation demand resting on all of them. The machine answers by mass: two thousand words of best practice, frameworks with names, a maturity model. Fluent, balanced, anchored to nothing in the business.
Sensing the answer is generic, the junior asks: is this the right approach for us? Criterion-free evaluation; the machine manufactures criteria from typicality and affirms, with caveats. The junior thinks aloud: maybe everything should just go through dispatch. Unmarked. The machine reads it as a premise, and the next thousand words design dispatch’s workflow. The session is now three levels below the question it began with, no return ticket, the upper levels never built. Around the second hour the junior senses wrongness and resets: step back, what are other ways to think about this? The reset discards the little that was load-bearing. By the third hour there is a transcript too long to reread, open questions multiplied, nothing decided, nothing written. Vapour. The junior is tired, and calls it discovery.
The second session. Same notes. The first message is different: these notes are rough; find the conflations and the unstated assumptions; ask one question at a time until the structure is sound; do not draft anything yet.
Pointed at structure rather than output, the machine surfaces in minutes what the first session never found. “Late” is undefined, and the notes use the word for four different things. No criteria exist for what better handling would mean. Zone-lead autonomy is mentioned, but its force is unclear: constraint, or preference? Each question comes back to the junior, who must answer from the one place the answers exist. Ten minutes of this is uncomfortable, and worth more than the previous session’s three hours.
Then the ladder, built by the junior, away from the machine. First a conceptual question: what counts as late, and what kinds exist. Then an empirical one: how comparable courier companies handle them, under an evidence contract. Then requirements, drafted as proposals and marked as proposals until the operations director confirms them. Only then design: two approaches, judged against the stated requirements. One excursion, declared: before comparing approaches, a probe three levels down to check whether rider-reported lates reach dispatch at all. They do not; dispatch learns only when a client calls. One variant of the central log dies on the spot. Ten minutes, for a week.
Checkpoints every few turns: decided, open, retired. Questions die as fast as they are born. Ninety minutes in, the session ends with a one-page document: the definition, the enumerated kinds, the requirements with their force marked, the chosen approach with its trade-offs stated, and three open questions that only the operations director can settle. Convergence, captured.
Same notes. Same model. Same junior. The model did not improve between sessions, and the junior did not become more talented. The variable was position: categories known, force marked, the ladder descended in order, every move declared. Everything else followed.
8. The Closing Inversion
The prevailing belief is that AI compresses the gap between junior and senior. For some it does. The machine is an amplifier of structure, and seniors are feeding it structure: their sessions begin with the question already well formed, descend in order, and end in artifacts. Their output has multiplied. The junior who feeds it confusion gets confusion back, at higher volume and in better prose. The gap does not close. It widens, and it widens precisely for the people who were promised it would close.
Two objections will be raised.
It will be said that better models will solve this. They will not, because the limit is informational, not technical. A model cannot recover intent that was never externalised; the structure of a particular problem is absent from the input, absent from the priors, and there is no third source. What a more capable model produces is a more convincing reconstruction of an underspecified intent. The failure does not become rarer. It becomes harder to see. On this one dimension, capability makes the problem worse.
It will be said that the protocol is slow. It is slow in the way seniors are slow: at the top, where an error costs a sentence, rather than at the bottom, where it costs the building. That arithmetic was settled in Section 3.
The deeper point is that nothing in this essay is new. Every word of the anatomy was true before the first model was trained. What has changed is that, for as long as work was done among people, colleagues quietly compensated for missing structure: the furrowed brow, the request for clarification, the interruption of the drift. The machine does not compensate. AI did not create the need for structured thinking. It removed the last place to hide from it.
So, finally, the loop. It will not be broken from above; the incentives are what they are, and essays do not repeal them. It can be broken at exactly one link: the quality of what the junior feeds the amplifier. Structure, unlike experience, does not take a decade to acquire. It is a vocabulary and a discipline: three axes, one law, four moves, a handful of habits. A junior who masters it becomes, in months, the person the machine multiplies, and walks out of the statistics used to justify the next round of cuts. One who does not becomes those statistics. The generation need not be lost. But it will not be rescued. The rungs that were pulled up will be rebuilt, where they are rebuilt at all, by the people standing where the ladder used to be.
Published by Shadow. Shadow is an AI operating system for PR and communications agencies.