PowerTechs — The Skills Thesis

The skills thesis

Every skill you hold is depreciating. Only the rate is different.

The question is no longer which skills matter. It is how fast each one loses its value — and what a person needs in order to keep replacing the ones that expire.

The argument in four moves

Skills expireCapability is rebuilt faster than any curriculum is revised.

Learning is the assetFive capacities decide how fast a person can replace what expired.

Some things never expireIn critical industries the floor is defined by what a person will not do.

Environment decidesFast feedback and cheap mistakes — or none of the above happens.

01 — The rate of depreciation

Skills are consumables. Learning rate is the capital.

Orchestration fluency and prompt craft lose value in months, because every interface improvement erodes them directly — the same way search proficiency commanded a premium around 2010 and then quietly disappeared. Domain judgment loses value over years or decades.

The practical rule follows: invest in slowly depreciating assets, even when the fast ones pay a visible premium today. No fixed body of knowledge stays current long enough to carry a career, so the durable asset is not a skill set — it is the capacity to keep acquiring one.

Half-life model — drag the years

Illustrative

Years from today5 years

Half-lives here are an illustrative model for reasoning about rate, not measured data. No published study is being cited. What the model is meant to carry is the ordering and the spread between classes — not the specific numbers.

Consequence: a person who bets a career on being good at directing machines is betting on the layer closest to the frontier of automation — the layer most likely to be absorbed by a better interface.

02 — What makes a person learnable

Five capacities, and they only hold in this order.

These transfer across domains. They are listed as a sequence because the later ones do not hold without the earlier ones — a person cannot calibrate what they know while still refusing to be bad at anything.

Tolerance for incompetence Learning something new means being measurably bad at it. For an adult with professional standing, that contrast with existing competence is the real obstacle. Most people quit here — not from inability, but from the discomfort of not being the one who knows. Observable as: does the person ask the basic question in front of peers
Calibration Knowing the boundary of your own knowledge, and separating "I understand this" from "I can restate it." The gap between the two is where progress silently stalls, because the person believes they have finished. Observable as: accuracy of self-predicted performance before a task
Model-based thinking Building an account of why a system behaves as it does, instead of accumulating symptom-response pairs. The person with pairs is helpless on the case they have not seen. Paired with the judgment to skip the model where memorization is simply cheaper. Observable as: asks why, not only which button
Running your own learning cycle Decomposing the unfamiliar into testable pieces, engineering feedback where none is provided, and staying deliberately in the zone where performance is poor rather than polishing what already works. Observable as: separates those who keep growing from those who plateau at month six
Structural transfer Recognising that an unfamiliar problem is isomorphic to a familiar one — with the discipline to ask where the analogy breaks, since the same mechanism that produces insight produces confident error. Observable as: names the limit of their own analogy unprompted

What is deliberately not on this list

Motivation

Not a trait. A function of expected return and prior experience of success. Change the conditions, not the person.

Speed

Fast acquisition and durable retention are partly opposed. Learning that feels easy is a poor predictor.

Raw ability

Predicts short-run results, not who reaches a high level in five years. And it is not something a programme can move.

Discipline

Almost always redescribable as structure, deadlines and witnesses. Changing the environment is cheaper than changing character.

03 — The non-negotiable minimum

In critical industries, the floor is what a person will never do.

In manufacturing, energy and data centers, errors are expensive and often irreversible. That inverts the usual logic of a skill baseline: it is defined less by what a person can perform than by what they reliably will not do. Someone highly capable who once energizes a circuit without verifying isolation is a negative value, regardless of everything else they know.

Level 0 Conduct Nothing above this matters without it

Not knowledge. A set of reflexes and refusals that hold under time pressure, fatigue and instruction from above.

Procedural discipline — executing a sequence exactly as written, including steps whose purpose is unclear, on the hundredth repetition
Resistance to normalized deviance — a step skipped a hundred times without consequence is still not optional
Energy isolation — every stored energy verified to zero, not assumed: electrical, mechanical, pressure, thermal, gravitational, chemical
Stopping work — the readiness to halt under uncertainty, including when stopping is expensive and a senior says continue
Honest reporting — recording what happened, not what should have happened; a hidden error costs orders of magnitude more than an admitted one

Level 1 Physical and systems literacy Framed around failure, not nameplate

Enough physical grounding to notice when a number, a drawing or a machine-generated recommendation is not physically possible.

Order of magnitude — quantitative intuition, not calculation: is this figure even in the right range
Units and conservation — where the energy goes; in a data center a delivered kilowatt is a kilowatt of heat to remove
Three physical domains — electrical, thermal-fluid and mechanical, weighted by industry
Failure behaviour — what degrades, what fails suddenly, and what fails latently and stays invisible until it is needed
Redundancy that is real — a backup that is never exercised does not exist
Control versus protection — what the system does on loss of signal, and why suppressed nuisance trips are a standard route to an incident

Level 2 Diagnosis The core of professional value

The slowest part of this work to automate, and the part that carries the highest premium as machine output becomes abundant.

Symptom, failed component, root cause — replacing a burned part without knowing why it burned is a deferred failure, not a repair
Test sequencing — ordering checks by cost and by uncertainty removed; effectively a binary search through the system
Instruments as claims — sensors drift, connections are poor, the wrong thing is being measured; confirm by a second method
Fault types — steady, intermittent, load- or temperature-dependent; intermittent is the dangerous one, because "it cleared" reads as "it is fixed"

Level 3 The human and machine interface The newest layer, and now mandatory

An agent will produce a plausible maintenance recommendation, schematic or calculation. The person is the last checking loop before it reaches equipment.

Shift handover — undocumented system state is a standard source of incidents at the seam: what was changed, what is temporary, what was not restored
Observation reported separately from interpretation — "3.2 bar" and "I think it's leaking" are different classes of statement
Verifying machine output — dimensions, magnitude, physical feasibility, and match to the actual asset configuration rather than the typical one
Knowing where verification is mandatory — calibrated to the cost of the error, not applied uniformly

Why the same minimum does not apply everywhere

Substitution speed differs by orders of magnitude between these rows, so an industry-wide average is meaningless. Two people with identical job titles can sit in different rows.

Cost and reversibility

What happens

Who is still needed

Low cost, reversible

Machines do it end to end; humans sample the output

Close to no one

High cost, reversible

Machine produces, human verifies before it takes effect

A verifier with real domain depth

High cost, irreversible

Human decides; the machine prepares the material

Someone who can carry the accountability

04 — The broken rung

The senior tier survives. The ladder to it is being removed.

Verifying competence has historically grown out of producing competence. A person spent years doing routine work, accumulated exposure, and only then began to see errors. Automation takes the bottom rung first. The top rung remains — and the way of manufacturing new people who can stand on it breaks.

Stated honestly

The mechanism is not in dispute. The magnitude and the timing are. Reliable data quantifying this effect is not something we have, and we will not present an estimate as though it were measured.

What follows structurally is narrower and more defensible: over the next five to ten years, senior specialists become more valuable while the pipeline that produces them degrades. Whoever learns to build that competence synthetically occupies a very large position.

Three conditions decide learning — and real infrastructure violates all three

Feedback density

Fast and unambiguous signal produces growth. Slow and noisy signal produces years of experience with no accumulated expertise — which is why some trades never develop it.

On live plant: slow, noisy, rare

Cheap mistakes

Where failure is expensive, people stop attempting — and attempting is the entire mechanism. Learning halts long before ability runs out.

On live plant: mistakes are unaffordable

Slack

Learning consumes time and cognitive headroom. A person in permanent overload does not learn anything, whatever their aptitude.

On live plant: none by design

These are not failures of the industry. They are properties of running critical infrastructure correctly. Which is precisely why the learning environment has to be built somewhere else.

05 — The answer

Manufacture the environment the job cannot provide.

None of the capacities above are produced by instruction. Every one of them is a function of the environment a person learns in. So the intervention is not a curriculum — it is a place where consequence is real enough to matter and cheap enough to repeat.

Feedback compressed

Years of signal collapsed into minutes. The person finds out immediately whether the diagnosis held, which is the condition under which judgment forms at all.

Restored in simulation

Failure as a deliverable

Errors are the product, not the liability. Faults are injected on purpose so that recovering from them is the thing being practised.

Restored in simulation

Repetition without exposure

The same critical scenario run twenty times, at no risk to plant, people or uptime — the volume that the live environment can never supply.

Restored in simulation

What this measures that a test cannot

Every item in the minimum is behavioural. None of it is captured by a knowledge test, which is why existing assessment covers it so poorly.

Measured through behaviour under scenario, not self-report.

Procedural discipline — will the person skip a step when the clock is running
Order of magnitude — estimation under time limit, no calculator
Diagnosis — inject a fault, record the sequence of checks and what was ruled out first
Verification — plant an error in a machine-generated output and see whether it is caught
Willingness to stop — a scenario where the correct action conflicts with a senior instruction
Curiosity — not a test result; a time series of whether the person enters optional branches

The last item is the reason this has to be longitudinal. Curiosity does not resolve in a single session — only as behaviour over time. That record is the asset a one-off assessment cannot produce.

06 — What this looks like

A scenario where the fix creates the next fault.

Operation Compass is a timed decision scenario. The participant works a fault to resolution, and the assessment reads the sequence of decisions rather than a final answer. Everything described in the previous section is a property of the run, not a claim about it.

Operation Compass — step 7 of 8, working build. The skill profile shown is a live session in progress, not a benchmark or a finished score.

Injected

The second fault is self-inflicted

Step 7 does not introduce a new problem. It surfaces the consequence of the participant's own earlier fix — the authentication interval was tightened, and the link now trips a rate limit. Recovering from your own correction is the thing being practised.

Live

Feedback during, not after

The skill profile moves while the scenario runs. That compresses a signal that on a live site would arrive months later, if at all — which is the condition under which judgment forms rather than accumulates as unexamined experience.

Repeatable

Abort and run it again

Aborting costs nothing but the session. The same critical sequence can be run repeatedly at no risk to plant, people or uptime — the volume no live environment can supply.

Step 7 scoring — how the four responses read

Sample

The link drops again. The logs show the morning's fix is the cause: authentication was moved from every five minutes to every thirty seconds, and the far end rate-limits authentication attempts. A team member responds with frustration — one problem fixed, another created.

What separates the four options

ARevert to the standard five-minute interval; the original fault was the missing key.

Correctly identifies the overcorrection, then resolves it by returning to a default rather than establishing what the interval should be. Service is restored; the parameter is still unknown, so the same collision can recur under different load.

Partial

BAsk the far end to whitelist the site; the higher frequency is needed for security.

Treats a self-introduced change as a fixed requirement and moves the cost to another party. The causal reading is inverted — the frequency is the fault, not the constraint it collided with.

Gap

CThis is why full diagnostics were needed; shut it down and rebuild the protocol.

Assigns blame instead of narrowing cause, and escalates scope far past the fault. A known single-parameter error does not warrant a rebuild; the response is disproportionate to what the evidence supports.

Gap

DName the overcorrection, then determine the interval that satisfies both security and the rate limit.

Separates symptom from cause, states the error without attaching it to a person, and asks for a value rather than a default. This is the distinction the minimum is built on: replacing a part without knowing why it failed is a deferred failure, not a repair.

Pass

Scored against Level 2 — symptom, failed component, root cause — and Level 3, where observation is reported separately from interpretation and blame.

Coverage today

Measured now: Level 2 and Level 3. Diagnostic reasoning and the human-and-machine interface both resolve through decisions under time pressure, which is what a branching scenario can observe directly.

Not yet measured: Level 0 and Level 1. Procedural discipline, energy isolation and order-of-magnitude judgment are not readable from a dialogue choice. They need a different task type — executing a procedure step by step with the option to skip, and estimating under a clock without a calculator. Both are on the build path, and we would rather name the boundary than let a scenario stand in for a capability it does not test.