The RFC Glut: What Happens to Design Review When Writing Is Free

A polished eight-page design doc used to mean something before anyone read a word of it. The existence of the artifact was evidence: someone spent two weeks thinking about failure modes, arguing with themselves about tradeoffs, and pre-empting the objections they knew reviewers would raise. The document was a receipt for the thinking. Reviewers could triage by polish alone, because polish was expensive to fake.

That correlation is now dead. An agent produces a comprehensive, well-structured, diagram-laden RFC in under an hour — complete with an alternatives-considered section for alternatives nobody actually considered. The artifact survived; the signal it carried did not. And most engineering organizations are still running a review process that was designed, implicitly, around the old cost of writing.

The result is a glut. Teams that once produced one RFC a month now produce five a week. Every one of them looks like it deserves a careful read, because looking like it deserves a careful read is precisely what language models are best at. Meanwhile the supply of review attention — senior engineers with the context to evaluate a design against reality — hasn't grown at all. Review attention is now the scarcest resource in the org, and it is being allocated by a mechanism that no longer works.

Design docs were a costly signal, and the cost just went to zero

Economists have a name for what design docs used to do: costly signaling. A signal works when it's expensive enough that only the genuinely committed bother to send it. The classic result is that the signal's information content comes from its cost, not its content — and when the cost collapses, the information collapses with it.

This isn't theoretical. A 2025 study of a large freelance labor market ("Making Talk Cheap," Galdin and Silbert) measured exactly this dynamic in job applications. Before LLMs, employers paid a premium for carefully customized proposals, because customization proxied for ability. After LLMs made customization free, that premium vanished — and the market got measurably less meritocratic. Top-quintile workers were hired 19% less often; bottom-quintile workers 14% more often. The employers didn't get dumber. Their screening instrument broke.

Swap "job application" for "design doc" and the mechanism transfers cleanly. A thorough RFC used to proxy for a thorough author. Reviewers leaned on that proxy constantly, even if they never articulated it: skim the doc, gauge the care, decide how much scrutiny to apply. When care is no longer required to produce the appearance of care, the skim tells you nothing. The reviewers didn't get dumber. Their instrument broke.

The workplace version of this already has a name: workslop. A 2025 Stanford and BetterUp survey of 1,150 U.S. workers found 40% had received AI-generated content that masqueraded as good work but lacked the substance to advance the task — and each instance cost the recipient nearly two hours of dealing-with-it time. That is the precise failure mode of a hollow RFC: it doesn't fail loudly, it fails by consuming the reviewer's afternoon before revealing there was nothing inside.

The bottleneck moved, and it moved to the most expensive people

The code-side version of this squeeze is well documented. Faros AI's telemetry across 10,000-plus developers found that high-AI-adoption teams merged 98% more pull requests — while review time rose 91% and PR size grew 154%. Generation scaled; judgment didn't. The same asymmetry hits design review harder, because design review was always more judgment-dense than code review. A PR can be partially validated by CI. An RFC is validated by nothing except a senior engineer's model of the system and the org.

So the glut lands directly on staff and principal engineers — the people whose attention was already the binding constraint. And it lands asymmetrically:

• Producing a doc takes an hour; seriously reviewing one still takes half a day. The cost ratio between writing and reviewing has inverted. Review used to be the cheap side of the transaction.
• Volume hides variance. When five docs arrive a week, some are genuinely thought-through and some are prompt-and-paste. They look identical. The reviewer must do deep work just to find out which kind they're holding.
• Refusing to read is politically expensive. Declining review reads as obstruction, especially when the doc looks complete. The old social contract — "I wrote eight careful pages, you owe me a careful read" — persists even though half its premise is gone.

The failure state isn't dramatic. It's quiet degradation: reviewers skim everything and deeply read nothing, approvals become rubber stamps, and design review stops catching the mistakes it exists to catch — while everyone continues performing the ceremony.

Move the signal to what documents can't fake

The fix is not banning AI-written docs. Drafting with a model is legitimately useful — it turns a blank page into an editable one and often surfaces considerations the author hadn't listed. The fix is to stop treating the document as evidence of thinking and start demanding evidence the document cannot fake. Three forms of it hold up:

Tradeoffs the author defends live. A model will happily generate an "alternatives considered" section, but it can't sit in a room and defend the choice against a skeptical staff engineer who asks why not the boring option? Shift the weight of review from reading to a short synchronous defense: thirty minutes, author presents nothing, reviewers ask questions the doc supposedly answers. An author who did the thinking handles this easily. An author who prompted a model into eight pages is exposed in the first ten minutes. The interrogation is the new costly signal — it can't be delegated.

Pre-registered success metrics. Prose commitments are now free, so demand falsifiable ones: what number will this design move, by how much, measured how, and by what date do we admit it failed? Borrowed from pre-registration in experimental science, this works because it converts writing into a bet. Bets are still costly. A generated doc will state benefits fluently; it takes an actual owner to stake a specific, checkable claim and accept being wrong in public.

Cheap prototypes over prose. When agents make code cheap too, a working spike is often no more expensive than the document describing it — and it's far harder to fake. "Here's a branch where the migration runs against a copy of prod data" dominates ten pages about the migration plan. The old rationale for docs-before-code was that code was the expensive artifact; where that has flipped, review the running thing and let the writing shrink to a record of the decision.

None of this makes documents worthless. It makes them what they always should have been: a communication medium, not a proof-of-work token.

Ration review bandwidth explicitly, or volume will ration it for you

Even with better signals, the arithmetic problem remains: doc production scaled 5x and review capacity scaled 0x. Any system where a free action (writing) compels a costly action (reviewing) from someone else will be driven to exhaustion — it's a resource-allocation bug, and it needs an explicit allocator instead of the implicit one ("whoever ships the most pages gets the most attention") that the glut just broke.

Concretely, that looks like:

• A review budget, stated out loud. Each senior reviewer commits to N deep reviews per cycle. When the queue exceeds the budget, docs wait or get triaged down — visibly, not by silently rotting in a queue that everyone pretends is being read.
• Tiered scrutiny by blast radius. A one-way-door decision — data model, public API, cross-team contract — gets the full treatment: live defense, pre-registered metrics, multiple reviewers. A reversible service-internal choice gets a lightweight async pass or none. The tier is declared up front by the author and is itself reviewable; misfiling a one-way door as reversible is the failure the process should punish hardest.
• A cost on submission. Require every RFC to open with three things written by a human: the single tradeoff that kept you up at night, the strongest argument against this design, and what you'll measure to know it failed. It's a small tax, but it restores a sliver of costly signaling at exactly the point where the reviewer decides whether to engage — and its absence is now informative.
• Author-side pre-review. Before a human reviewer engages, the author runs an adversarial pass — with a model, fine — and attaches the objections it raised plus their responses. This flips the same cheap generation that caused the glut into a filter against it: if a machine can find the hole in ten minutes, a principal engineer shouldn't have to.

Notice what these have in common: they all price attention. The pre-AI RFC process never had to, because the cost of writing rationed demand automatically. That subsidy is gone.

The signal always migrates — follow it

There's a general pattern here worth naming, because RFCs are only the first place it's biting. Whenever an artifact is used as a proxy for effort — the cover letter, the take-home assignment, the quarterly strategy memo, the design doc — cheap generation kills the proxy while leaving the artifact standing. Organizations then face a choice: keep performing evaluation rituals around a dead signal, or find where the signal moved and rebuild the process around its new location.

The signal has moved to whatever still can't be generated: live defense under questioning, staked predictions that resolve true or false, working systems under load, and track records of past bets that paid off. Each of these is costly in a way that fluent text no longer is — which is exactly why they're now the only things worth reviewing.

Design review isn't dying; it's being repriced. The organizations that come out ahead won't be the ones producing the most impressive documents — that capability is now table stakes, available to everyone for the cost of a prompt. They'll be the ones that noticed their review process was quietly reading receipts for thinking that no longer prove any thinking occurred, and moved their scarce attention to the evidence that still does.