The Hidden Labour of Externalizing Expertise
The Work Beneath the Prompt
The public vocabulary for human-machine work remains too thin. We say that someone “used AI,” “wrote a prompt,” or “generated a draft.” These phrases describe visible actions, but they miss much of the work that determines whether a machine-assisted output is useful, weak, misleading, or genuinely valuable.
In serious knowledge work, the decisive human contribution often happens before the visible output appears. It lies in the expert’s ability to convert purpose, judgment, standards, and tacit expectations into forms a machine can act on. This is the hidden labour of externalizing expertise.
Experts rarely operate by applying fully explicit rules. Much of their competence is compressed into habits of attention. They know which details matter, which claims sound plausible but weak, which categories are too loose, which sources are admissible, which patterns are meaningful, and which outputs are fluent nonsense. They know when a result is technically correct but strategically useless. They know when a draft has the right information but the wrong posture. They know when to stop, when to push, when to discard, and when to start again.
Human collaborators often absorb these standards through apprenticeship, shared disciplinary background, repeated correction, and institutional practice. Machine agents do not absorb them in the same way within a project. They require articulation. The expert must surface what normally remains tacit, then translate it into instructions, constraints, examples, schemas, rubrics, prompts, file structures, and correction loops.
That labour is easily mistaken for simple prompting because the interface reduces everything to text in a box. The interface makes expert orchestration look like casual instruction. This is a serious distortion. A weak user may ask for a report. An expert orchestrator defines the purpose of the report, the allowable evidence, the structure of the argument, the expected level of inference, the rhetorical stance, the technical constraints, the failure modes to avoid, and the standards by which the output will be judged.
The difference is not cosmetic. It is the difference between assignment and governance.
Externalization as Contribution
Externalization is often treated as a neutral conversion process: an idea becomes a note, a note becomes a plan, a plan becomes a document. In human-machine work, externalization becomes more consequential. It is the process through which human expertise becomes machine-operable.
A practical sequence can be stated simply: tacit judgment becomes an articulated heuristic; the heuristic becomes an operational rule; the rule becomes a protocol; the protocol becomes a schema; the schema becomes part of a workflow; the workflow produces an auditable trace.
That sequence is rarely clean. It usually involves trial, failure, revision, and correction. The expert discovers that what seemed obvious in practice was never actually explicit. The machine fails because the instruction was ambiguous. The schema fails because the category boundary was too vague. The prompt fails because it requested output without specifying standards. The workflow fails because it assumed that a fluent response was the same thing as a valid one.
This is where serious orchestration begins. The expert must observe failure closely enough to see which part of the tacit system was missing. Was the task poorly decomposed? Was the source hierarchy unclear? Was the model given too much context or too little? Was the requested output format underspecified? Was the evaluative standard absent? Was the machine asked to make a judgment that should have remained with the human? Each correction externalizes another piece of the expert’s working intelligence.
This process has a long ancestry in attempts to formalize expert knowledge. What is different now is the practical setting. Earlier expert systems tried to encode knowledge into explicit rules. Current machine agents already generate, summarize, classify, draft, and reason across large bodies of material. The main problem is often less that the machine cannot produce anything and more that it produces too easily. It needs constraints, standards, roles, stopping rules, quality checks, and forms of accountability. The expert is no longer only transferring knowledge into a system. The expert is building conditions under which machine participation becomes admissible.
That is a real contribution. It may leave fewer visible traces than a paragraph of prose or a line of code, but it can shape the entire output. A model may generate a draft, yet the draft may depend on a human-designed task contract, a structured prompt, a curated context window, a source hierarchy, a validation protocol, and repeated rejection of inferior responses. To attribute the result only to the machine that generated the words, or only to the human whose name appears on the page, is to misrepresent the activity system that produced the work.
The hidden contribution is especially important because it is cumulative. Once externalized, expert judgment can be reused. A strong schema can discipline future outputs. A good prompt can become a task harness. A correction pattern can become an evaluation rule. A failed run can become a test case. A workflow can stabilize a practice that previously depended on memory, improvisation, and individual tolerance for confusion.
Externalization converts private competence into shared infrastructure.
Beyond Casual Prompting
The distinction between casual prompting and expert-led orchestration matters because institutions will be tempted to flatten both into the same category. That would be a mistake.
Casual prompting delegates a task. Expert-led orchestration configures a task environment. Casual prompting asks for an answer. Expert-led orchestration specifies what kind of answer can count, under what conditions, using which materials, with which exclusions, and subject to what tests. Casual prompting accepts or rejects the surface output. Expert-led orchestration examines how the output was produced and whether the process can be repeated, audited, corrected, and improved.
This difference has consequences for attribution. A person who writes a short instruction and accepts the result has contributed differently from a person who designs the workflow, externalizes disciplinary standards, iterates correction loops, validates outputs, and carries responsibility for publication. Both may appear to have “prompted the AI.” Only one has performed expert orchestration.
This distinction also has consequences for training. The future of serious knowledge work will not be secured by teaching people a few prompt formulas. The harder task is teaching experts to recognize their own tacit standards and express them in forms that can guide machines. That requires reflection, observation, reduction, abstraction, and testing. It requires the expert to move between human-facing expression and machine-actionable structure. It requires comfort with prose, tables, markdown, JSON, scripts, logs, and versioned artifacts. It also requires enough disciplinary confidence to push back against machine fluency.
Machine systems can produce plausible outputs at speed. The expert must supply resistance. That resistance is not hostility to the machine. It is the imposition of standards. It is the refusal to confuse surface coherence with adequate work.
This is why externalization labour should be visible in any serious attribution system for human-machine projects. The contribution is not merely that the expert had an idea. It is that the expert transformed situated judgment into operational conditions for machine work. That transformation may shape the whole project: what gets asked, what gets excluded, what counts as evidence, what is considered failure, how revision proceeds, and when the output becomes acceptable.
The prompt is only the visible edge of this labour. Beneath it is a larger act of translation: from tacit expertise to structured participation, from private judgment to shared control, from human habit to machine-operable workflow.
If human-machine work is becoming a new division of labour, externalization is one of its central acts. It is also one of the easiest to overlook. The next attribution systems should make it visible.