Why this is in the vault
Every senior editor Jack Cheng argues that hitting AI usage limits is a forcing function for better work — the wait lets his best questions develop and exposes when he's "snacking" (cheap, low-value tasks) instead of doing hard work — which is a direct outside data point on RDCO's own model-effort pairing and no-babysitting discipline.
The core argument
Cheng used to exhaust his weekly Claude/Fable allocation in a day, then fall back to faster, lesser models (Sonnet, Sol) for routine work while saving his hardest questions and one-shot builds for when his "oracle" limits reset — a ritual he compares to consulting the oracle at Delphi, where the journey itself, not just the answer, clarified the question. As frontier models (Fable 5.1) got more token-efficient, his weekly budget started stretching three to four days, and he noticed he wasn't better off for it: he started reflexively sending every feature idea to Fable without pausing to ask whether the work was worth doing, engaging in what Will Larson calls "snacking" — choosing easy, low-impact work over difficult, high-impact work. He traces the same tension through computing history (mainframe queues, time-sharing, Alan Kay's "object to think with" becoming an "object to think for") and lands on a practical fix: deliberately reintroduce friction — let a feature idea sit overnight, write down why a question matters and what he thinks the answer is before bringing it to a model — without actually downgrading his plan.
Mapping against Ray Data Co
Cheng's "snacking" diagnosis is the same failure mode the feedback_no_babysitting and feedback_studio_throughput_no_deferral memories already guard against from the opposite direction: RDCO's answer to unlimited apparent capacity is not to throttle access but to route effort deliberately — feedback_delegation_model_effort_pairing already mandates that Fable delegations get an explicit high/xhigh effort setting reserved for "meaningfully large tasks only," which is functionally Cheng's own fix (force a value judgment before spending capacity) without needing an artificial scarcity signal. Where Cheng leans on quota exhaustion as an external discipline mechanism, RDCO already tries to build the discipline into the dispatch decision itself — this essay is a useful outside check that the underlying risk (cheap capacity erodes the habit of asking "does this deserve to be built") is real and not just a hypothetical RDCO is over-engineering against.
The essay's "let it sit overnight" fix also lands directly on the founder's own inflection point tracked in user_money_values_potential_tension — deliberately creating space between an impulse and an action so the impulse can be evaluated rather than immediately executed, aimed at spending/build decisions here rather than money, but the same underlying mechanism.
Sponsorship
No paid sponsor block found in this issue. The footer carries Every's standard house self-promotion (Every All Access membership + its own product bundle: Sparkle, Cora, Spiral, Monologue) — boilerplate present in nearly every Every issue, not a one-off paid or rotating sponsor. Marked sponsored: false accordingly.
Related
- [[2026-01-28-every-stop-coding-start-planning]]
- [[2026-08-17-every-ai-costs-token-budgets]]
- [[feedback_delegation_model_effort_pairing]]
- [[feedback_no_babysitting]]