What it was for
The PRD served two masters: a persuasion document that won stakeholders, and a translation document that told engineers what to build. It was written for human readers, judged by its narrative, and routinely outdated before the build it described had finished.
The verdict
TRANSFORMS. In the engine, the PRD's translation function is absorbed into intent specification — and the standard changes from convincing to executable. A requirement now has a new and unforgiving reader: a machine that will do exactly what the document says, at speed, including the ambiguities. Prose that a charitable engineer would have quietly repaired, the inner loop will faithfully ship.
What changes
The artifact splits along its two old jobs. The persuasion half remains human writing — context, stakes, the why — and gets shorter and more honest. The translation half hardens into machined intent: acceptance criteria as evaluations, constraints stated as rules, edge cases enumerated rather than implied. "The team will figure it out" stops being a sentence a PM can write, because the team that figures it out now includes machines that figure it out wrong, confidently.
The strongest objection
Over-specification was the disease waterfall died of; this sounds like its return. Conceded as the failure mode to watch — but the cure differs from the disease. Waterfall specified everything once, upfront, blind. The engine specifies one intent per revolution, against evidence, continuously. Precision per loop, not prophecy per year.
Corroboration (Jul 2026). The Thoughtworks retreat found "traditional user stories are too vague" for agents and teams reaching for structured formats — EARS, state machines, decision tables — to make intent precise enough that a machine cannot confidently build the wrong thing. That is this entry's shift of standard from convincing to executable. Ng's parallel: engineers absorbing product work on the strength of "context advantage." (Thoughtworks retreat; Andrew Ng · The Batch #359)
Corroboration — the spec as the only artifact (Sep 2026)
The strongest version of this entry's claim is now a working library rather than an argument. Design Docs Are All You Need (Kushnir et al., September 2026, from MIT, Google, Stanford and Google DeepMind) describes smart, a symbolic ML performance-modeling library whose main branch "contains almost no code." The repository is a DAG of 50 natural-language design docs, roughly 9,000 lines of specification prose, and coding sub-agents regenerate the entire implementation from them. "Humans only ever edit the docs."
A clean-slate regeneration takes between 1.5 and 3 hours and costs around 100 USD, which the authors put at roughly 20% of a weekly usage budget. The regenerated implementation reproduces hand-audited reference models to round-off precision. They used it to model DeepSeek-V3 serving on a TPU pod slice.
That is the translation half of this entry taken to its limit — the document stops describing the build and becomes it.
The condition it runs on
Regeneration is safe there because the verification leg is unusually strong, and the paper never says so. Every rebuild "must pass reconciliation against hand-built references before it replaces the previous build," and "every number-bearing doc ends with a reconciliation anchor: a small preset whose expected outputs are stated exactly and enforced by generated tests." They have a hand-audited oracle and exact expected numbers, on a problem where correctness is checkable to round-off precision. Most software has no such oracle, and the paper has no limitations section and never asks which domains this suits.
So it demonstrates something narrower than its title and more useful: when specification and verification are both strong, throughput becomes disposable. That is evidence for the two legs together rather than for documents alone — which is also the warning. Copy the doc-as-source-of-truth structure into a domain with no reconciliation anchor, and what regenerates is confident, fluent and unchecked.
One correction worth carrying if you cite it. "Almost no code" is the title's rhetoric rather than the result. Nine thousand lines of specification is not an absent artifact, it is an artifact in a different language, which is what TRANSFORMS has said here since July.