# Cheap Generation Raises the Value of Deterministic Verification ## Claim As AI makes code and interface generation cheap, the scarce and valuable work shifts to deterministic verification — tests, linters, hooks, and decision records — because prose instructions can't reliably carry repeatable constraints the way a check can. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds specifically for *repeatable* constraints — formatting rules, regression behavior, architectural decisions — that a deterministic system can enforce every time without re-deriving them from context. - Doesn't hold for constraints that are genuinely judgment calls each time (does this interaction feel right, is this the correct product tradeoff) — those stay human work regardless of how cheap generation gets. - Applies most clearly in agentic-coding contexts where an agent re-reads its instruction context on every task; a human developer with durable memory of team norms needs the deterministic system less urgently for the same constraints. ## Evidence Syntax's design-engineering episode argues that once AI can imitate visual conventions, interaction quality and usability become the real differentiator, and that coding-agent efficiency comes from moving repeatable rules out of prose and into tests, linters, hooks, and ADRs: "make it as deterministic as possible." ## Links - [[AI & the Future of Knowledge Work MOC]] — the core reallocation this MOC tracks — generation got cheap, checking did not - [[3. Resources/Permanent Notes/External Test Oracles Make AI-Generated Code Trustworthy]] — because evidence-for; an independently-arrived-at instance of the same mechanism from a different source (a from-scratch PHP interpreter validated against an external oracle it didn't author) — both converge on "a check you didn't write yourself is what makes generated output trustworthy." - [[3. Resources/Permanent Notes/Decision Quality Is the Competitive Variable in AI-Augmented Builds]] — because extends; that note locates the scarce resource in a builder's upstream decisions (problem, user, scope), this note locates a second scarce resource downstream, in verification — together they bracket where human judgment remains load-bearing once generation is cheap. - [[3. Resources/Permanent Notes/Deterministic Skill Steps Belong in Scripts Not the Model]] — because extends; this note moves verification out of prose into checks, that one moves execution of one-right-answer steps out of the model into scripts (Penn, 2026-10-04). - [[3. Resources/Permanent Notes/A Process Whose Value Is the Thinking Cannot Be Automated by Its Output]] — because applies-in; deterministic checks are exactly the kind of artifact that can absorb repeatable thinking without losing it, unlike prose instructions an agent has to re-interpret each time. - [[3. Resources/Readwise/Articles/Why Everyone Is Moving to Stylex]] — because evidence-for; a separate source reaching the same conclusion from CSS specifically — StyleX's typed atomic classes turn "does this match the design system" from a prose suggestion into a compile error, and the episode's own strongest line names the mechanism directly: "when you have a rigid way where an agent or yourself can know is this allowed or not, then it is much easier to stay within the lines." The unresolved instance it raises is WordPress's theme.json, which defines design tokens but enforces nothing on block styles or inline CSS. ## Tests / falsifiers - Compare agent task outcomes with instructions given only as prose context versus the same constraints expressed as tests/linters/hooks; if error rates and token cost don't improve with the deterministic version, the claim's mechanism is wrong. - If teams report that agent instruction files stay small and stable over time (rather than growing as more edge cases get added in prose), that's evidence constraints are actually migrating into deterministic systems as this claim predicts. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/The Rise of the Design Engineer]] — Syntax's design-engineering and coding-agent-efficiency discussion.