Module 03 · signal fire · Specification ·

Say What You Mean, Precisely

Execution became cheap; transferring intent did not. Communication turns into specification, and specification is a learnable craft.

Projection · scene

Lyon, 2036. The junior architect’s agent delivered exactly what she asked for: a bridge design minimizing material cost under the loading code. It is beautiful, compliant, and unbuildable — the erection sequence requires a crane that cannot stand in the river. Her senior looks at it the way seniors have looked at such things for a century: “You didn’t say it had to be constructible. You never have to say that to a person.”

The senior’s actual job, she realizes, has become knowing everything that people never had to say — and saying it.

The mechanism

Human collaboration runs on a vast unstated commons: shared context, professional defaults, the thousand constraints “any reasonable person” fills in. Delegation to machines drains that commons. An optimizing agent does exactly what the objective says — that is what optimizing means — and every constraint you leave unstated is, formally, a constraint you have priced at zero. The agent is neither malicious nor stupid; your specification really is the only channel your intent travels through.

This was always true of software; programmers just absorbed the pain professionally. Under baseline B1, delegation becomes everyone’s daily act, so the pain generalizes: the gap between what you know and what you state becomes the main place where value leaks. And the two curves compound: as execution gets cheaper and more literal, each unit of specification skill moves more output. The economists’ word for that is appreciation.

Execution is no longer the bottleneck; intent transfer is. The expensive sentence is the one you didn’t know you had to say.

The craft is learnable because it always existed: it is what good managers, good clients, and good senior engineers were doing all along — stating acceptance criteria, naming the failure modes, declaring what must not be optimized away. The decade just promotes it from soft skill to production function.

⏵ Toy 03 · The literal-minded agent

task: “arrange the team offsite” · spec bandwidth: 3 constraints — the agent optimizes exactly what you state, and nothing you don’t

Brief the offsite agent: eight constraints you know about, bandwidth to state three. The agent satisfies your words perfectly (src/lib/specification/, tested). The complaints write themselves.
Prepare now · 2026 · coals to carry
  1. Next delegation — human or machine — write the acceptance test first: “I will consider this done when ___.” If you cannot finish the sentence, the task is not ready to hand off.
  2. Collect your own violated-constraint list: each time output surprises you, record the constraint you assumed but never stated. Three months of that list is a specification curriculum tailored to you.
  3. Practice the negative spec: for one real task, write down what must not change, break, or be traded away. Unstated invariants are where literal-minded optimizers do their damage.