Technical writer / documentation engineer

Semi-structured interviews with 31 experienced technical writers at startups, enterprises and open-source projects. The authors describe five review stages: 28 participants described technical review by engineers or other subject experts, which surfaces knowledge not written down elsewhere; 13 described play testing, where people with product knowledge who did not write the documentation follow it end to end to find what the writer is too close to see, mostly for tutorials; and 25 described LLM-generated review feedback as unreliable, partly because the context a review needs sits in confidential internal systems they cannot share with model providers. Writers report they hold no authority to compel review and must negotiate for it. It is a qualitative study of self-reported practice; one author works at a company that sells language models.

Measured, large-scale use of a tool for real work, where the decision to use it was the worker's rather than an employer's. It is more than a capability record — the work is real, not a demo — and less than a deployment record, because no employer put it into production, required it, or built a process around it. Weighted `cautious`: `automating` means the machine can do the task AND there are adoption signs, and this is an adoption sign — but usage can be experimental, and much of the measurement comes from a party with a stake, so one record is never enough and two independent ones are. Note who is counting. Vendor telemetry sees this directly and sells the tool, so such a record names that stake in its scope; a statistics agency asking firms whether their workers use AI in tasks sees the same channel with no stake at all, and that is the better source where it exists.

FDA's guidance for medical device makers, marked as containing nonbinding recommendations. It covers labelling including instructions for use, user manuals, package inserts and quick-start guides, and says human factors validation testing can assess the adequacy of the instructions for use, but only in the context of use of the device; that the most important consideration for test participants is that they represent the population of intended users; and that if testing shows use errors on critical tasks, stating that the risk was mitigated by modifying the instructions or other labelling is not acceptable without additional test data showing the change worked. It applies to medical devices only and does not say who writes the instructions.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

Read article by article on the Ministry of Government Legislation's portal, in the version in force from 21 July 2026. Two sub-paragraphs of Article 34(1) bear on this occupation and they do different things. Sub-paragraph 2 requires the operator to establish and implement a plan for explaining — so far as technically possible — the final result the AI produced, the main criteria used to reach it, and an outline of the training data used to develop and operate it. Sub-paragraph 5 requires the writing and retention of documents by which the content of the safety and reliability measures can be verified. The first governs what a reference has to cover; the second removes the discretion to leave something out. Article 36(1)(3) then makes the point about maintenance rather than authorship: a foreign operator above user or revenue thresholds must appoint a domestic representative whose duties include supporting compliance with Article 34(1), expressly including checking those documents for currency and accuracy. Two limits. The act leaves the detail to Presidential Decree and to a ministerial notice, and Article 34(2) says the minister may recommend compliance with that notice — so the specificity of what must be written is not fixed by this record. And it binds only operators of high-impact AI in Korea; nothing here counts a technical writer, a document or an hour.

Regulation, subsidy or public procurement is requiring or funding adoption — the mirror of a constraint. It shows adoption is being required, not that it has happened, so one mandate is never enough on its own; two independent ones are.

Nationally representative, reference period November 2025 to January 2026. This is the worker-task layer of the same supplement rather than its firm-adoption layer: 23% of firms (41% employment-weighted) report workers using AI in work-related tasks, and among the tasks reported, writing, document analysis and information search lead. It is a count of firms reporting a task, not a measure of how much writing is done that way, and it does not distinguish technical documentation from any other writing — nobody asked what was written. Its value here is what it is not: unlike vendor telemetry, this is a statistics agency counting the same channel with no stake in the answer, which makes it the better source for worker adoption where it exists.

Measured, large-scale use of a tool for real work, where the decision to use it was the worker's rather than an employer's. It is more than a capability record — the work is real, not a demo — and less than a deployment record, because no employer put it into production, required it, or built a process around it. Weighted `cautious`: `automating` means the machine can do the task AND there are adoption signs, and this is an adoption sign — but usage can be experimental, and much of the measurement comes from a party with a stake, so one record is never enough and two independent ones are. Note who is counting. Vendor telemetry sees this directly and sells the tool, so such a record names that stake in its scope; a statistics agency asking firms whether their workers use AI in tasks sees the same channel with no stake at all, and that is the better source where it exists.

The task that used to hire you — producing the reference — is the most exposed writing task on this site. The one that is now scarce is the one nobody hires for directly: being the person who does not know, follows the instructions, and finds the step that is impossible. Make that visible in how you work, because it is the only argument that survives cheap drafting.

Your team will be measured on coverage and coverage has just become free, so the metric will look excellent while the docs get worse in a way nobody can see from a dashboard. The number worth putting in front of a manager is support tickets on the top ten pages, because that is the only measure that moves when the documentation is actually good.

Reshape the role

Become the first user, formally

The only documentation input a model cannot have is the experience of being confused, and a team shipping fast needs somebody whose job is to be confused on purpose before customers are.

Adjacent move

Move to documentation that is a compliance artefact

In regulated fields the document is required by rule — Korea's AI act makes high-impact operators write and retain documents evidencing their safety measures — so it is held in place by law rather than by budget.

Drafting the reference is the most exposed writing task on this site, because the source of truth is machine-readable and the output format is fixed — and doc generators had already taken part of it long before models. What does not move is the test: following your own instructions on a clean machine and finding the step that is impossible for everyone who is not the engineer. A model trained on the codebase inherits exactly the knowledge that has to be absent for that test to work.

No date. The signal is what your team is measured on. If it is pages published or coverage, the metric has just become free and will look excellent while the docs get worse — and headcount decisions get made off that dashboard. Watch whether anyone is tracking support tickets against the top pages, because that is the only measure that moves when documentation is actually good, and whether it exists tells you how the next budget will go.

It stopped being one the moment generation became cheap, and that is the most consequential change in this occupation. Coverage now measures how much text exists rather than whether anybody can do anything with it, so a team judged on coverage will be rewarded for abandoning the task that just became the important one — deciding what not to write. If your team still reports coverage, raising this is more valuable than writing another page.

Materially, and it is the clearest example on this page of a task held in place by something other than cost. In medical devices the instructions are judged by how representative users perform with the device, and FDA guidance says fixing a use problem by rewriting them is not acceptable without new test data, which means the work cannot simply be generated and shipped. That is a policy protection rather than a technical one — durable exactly as long as the rule stays — but it is real today.