"Sometimes You Have to Delete Everything" — Every Staff / Context Window (Jul 26 2026)

Why this is in the vault

This is Every's weekly internal digest ("Context Window"), bundling links to the week's Every output plus one original closing essay by Ashwin Sharma on why copying an expert's prompts doesn't confer their expertise. The essay is the load-bearing content; the rest of the issue is curation of Every's own week (Opus 5 Vibe Check, an All Access launch retro, a workflow-retention post-mortem, a Whoop/Maven partnered piece) plus house self-promo.

The core argument

Sharma watched mathematician Terence Tao use ChatGPT to explore a counterexample to the Jacobian conjecture and noticed Tao's prompts were short and unremarkable — the value came entirely from Tao's ability to judge whether the model's answers were correct, not from prompt cleverness. This undercuts the popular idea that copying an expert's prompts transfers their expertise. He pairs it with an Anthropic study in which AI-assisted junior developers scored worse on a post-task quiz than developers who coded by hand (50% vs. 67%), especially on debugging — precisely the skill needed to catch AI errors. His conclusion: easy AI answers let people skip the "stuckness" that builds real expertise, so AI likely makes genuine expertise more valuable while making experts harder to produce. Two lines that carry the argument: "expertise is knowing when it hasn't" and "AI will make expertise more valuable and experts harder to produce."

Issue contents

Beyond the closing essay, this issue links to (all Every's own published pieces, curated by Kate Lee):

"Claude Opus 5 Is Brilliant in Flashes, Frustrating in Practice" (Dan Shipper / Katie Parrott, Vibe Check) — filed separately at [[2026-07-24-every-vibe-check-opus5]].
"How Every's Team Used AI to Ship Its Biggest Launch Ever" (Laura Entis, Context Window) — All Access launch retro; Every COO Brandon Gell interviews the growth/marketing team behind a ~$9,000-in-two-days launch. Also mentions OpenAI's Codex playbook and a Fable token-cost-cutting technique via sub-agent delegation.
"Why Some AI Workflows Stick—And Others Don't" (Katie Parrott, Working Overtime) — a post-mortem on an abandoned internal tool, framed as four questions for deciding whether a workflow survives on what it costs vs. what it gives back.
"Drowning in Demos? Here's a Better Way to Prototype" (Hilary Gridley, Whoop) — explicitly marked as produced in partnership with Maven; argues that once building is cheap, a prototype's job is to test whether the problem is worth solving, validated by users not stakeholders.
"Steal this workflow: How Notion builds Notion with AI" — Notion engineer Ryan Nystrom's voice-first coding workflow: talk through the task, hand it to an agent, run a custom review swarm before human review.

No third-party curation beyond Every's own verticals — the "external" links inside the Sharma essay (Tao's ChatGPT share link, an Anthropic research page) are citations within Every's own essay, not a separate curated links block.

⚠️ Sponsorship

House self-promo, two distinct instances, both disclosed here:

Every All Access / Builder Pack upsell appears twice — mid-issue ("Upgrade to Every All Access and get six months of Notion Business... redeem more than $7,000 in partner offers") and in the footer paywall block, which also promotes Every's own products (Sparkle, Cora, Spiral, Monologue) and a "From Every Studio" section previewing Cora's relaunch and a new "Every Agent" Slack coworker in alpha.
Paid partnership disclosure within curated content: the Hilary Gridley/Whoop prototyping piece is explicitly marked "(This piece was produced in partnership with Maven)" and plugs her paid Maven course with a discount code. Treat any positive framing of that piece as commercially incentivized, not neutral curation.

Neither instance disqualifies the issue — the closing essay is independent editorial — but both should inform how much weight to give the "Knowledge base" summaries as neutral judgment.

Mapping against Ray Data Co

Direct hit on the DSA/TAL cert push and the founder's build-then-teach commitment. Sharma's finding — expertise is the ability to evaluate an AI's answer, not the ability to write a clever prompt — is the exact skill the phData Snowflake GenAI and Anthropic Architect certs are meant to prove founder-side, and it's the same skill the RDCO harness assumes Ray supplies when agents report back "DECISION:" lines for founder judgment (per feedback_advisor_not_pair_programmer and the verify-* fresh-eyes gate family). If the founder ever lets agent output substitute for his own evaluation loop rather than train it, this essay says that's exactly the trap — the Anthropic junior-developer study found AI assistance made debugging skill worse, not neutral. It argues for treating every agent-produced artifact as a chance to sharpen judgment, not just an output to approve.

Secondary connection: the issue's other pieces (workflow-retention post-mortem, prototype-validation-by-users-not-stakeholders) both reinforce the standing RDCO practice of routing verification through independent fresh-eyes subagents rather than the producing agent's own confidence — Sharma's "stuckness builds expertise" argument is the human-side mirror of that same principle.

[[2026-07-24-every-vibe-check-opus5]] — same issue's own link, same Every publication, same week; the Vibe Check this digest summarizes in one paragraph
[[2026-07-25-every-opus5-compound-engineering-breakage]] — the primary-source deep-dive on the Opus 5 breakage this digest references in passing; that note's mapping section is the more load-bearing RDCO hit on the model-migration side
[[feedback_advisor_not_pair_programmer]] — the founder-as-evaluator role this essay's core argument directly validates
[[feedback_fresh_eyes_subagent_for_own_artifacts]] — the RDCO practice of routing verification through independent judgment rather than the producing agent, structurally parallel to Sharma's "expertise is knowing when it hasn't" claim