"How We Work Now" — Every Staff

Why this is in the vault

A Sunday Context Window roundup curating Every's week of AI-native work pieces (Codex-for-knowledge-work guide, compound-engineering v2, an Opus 4.8 Vibe Check, and a Utah AI-prescribing pilot), several of which directly inform RDCO's agent-harness and operating-rhythm thinking.

Issue contents

Curation issue. Per-item provenance noted (third-party vs Every self-cross-promo):

1. "Codex for Knowledge Work" by Katie Parrott (Every — Guides). A ~9,000-word guide reframing Codex as an operating system for knowledge work: five levels of use (one-off tasks → compounding systems), 13 workflow templates, and the context-file / rules / review-checklist setup that makes agents reliable across a full workday. Companion framing essay for newcomers. Every self-cross-promo (own product/guide).

2. "Compound Engineering" by Kieran Klaassen and Trevin Chow (Every — Guides). The compound-engineering loop expands from four steps to seven; ideate and plan move to the front and polish to the end now that AI handles the middle. Updated plugin ships 43 subagents and 38 slash-command skills. Companion essay frames it as a "sandwich" — AI in the middle, humans as the bread on either end. Every self-cross-promo.

3. "Vibe Check: Opus 4.8 — Anthropic Should've Rounded Up to 5" by Dan Shipper and Katie Parrott (Every — Vibe Check). First Anthropic release in a year the authors would reach for across coding, prose, and everyday work. Scored 63 on Every's Senior Engineer Benchmark (vs 62 for GPT-5.5, 33.5 for Opus 4.7) and 79.6 on writing tests — highest any model has hit, with fewer "AI tells" than any non-Claude model. Verdict: the model now outpaces the app built around it. Every editorial (model review).

4. "We Automated Everything With AI and Tripled Our Headcount" by Dan Shipper (Every — AI & I podcast). Audio/video discussion of Dan's "After Automation" thesis: better models create more work for humans, not less (more frames to hand them). Every COO Brandon Gell presses on each premise. Every self-cross-promo (own podcast + own essay).

5. "After 'After Automation'" by Katie Parrott (Every — Context Window). Reads Pope Leo XIV's Magnifica Humanitas — described as the Vatican's first major AI encyclical — as a collective companion to Dan's labor thesis. Every editorial.

6. Camps / events block (Every). Compound Engineering Camp (June 5), Codex Camp (June 12), Executive AI Sessions webinar via Every Consulting (June 2), and a subscriber IRL meetup in Brooklyn (June 3). Every self-cross-promo (events + consulting funnel).

7. "From Every Studio: Proof keeps your name on shared docs" (Every). Product update for Proof (human + AI agents co-editing docs): eight PRs on collaborative editing, edits now attributed to the first human who opens a shared doc. Every self-cross-promo (own product).

8. "Alignment: The right kind of nervous" by Ashwin Sharma (Every). Commentary on the first five months of Doctronic's Utah AI prescription-renewal pilot (Office of AI Policy results). AI recommended renewal in 72% of cases; reviewing physician agreed ~90% of the time; after second-physician review 97% of recommendations stood; the office estimates human error at 5–12%. Of the 28% escalated to a physician, doctors backed the AI's caution 69% of the time. Sharma's frame: the system is behaving like a "nervous junior" doctor, and at this stage that overcaution is the encouraging signal. Every editorial, citing a third-party government report.

9. Footer product line + paid upgrade CTA (Every). Spiral, Sparkle, Cora, Monologue, Proof bundle pitch. Every self-cross-promo. A "for sponsorship opportunities, reach out to sponsorships@every.to" solicitation appears but no paid sponsor block is present in this issue.

Mapping against Ray Data Co

Mapping strength: STRONG. Multiple items map onto active RDCO themes; the Opus 4.8 item is directly load-bearing.

• Opus 4.8 Vibe Check → this session's substrate. This very COO session runs Opus 4.8. Every's independent benchmark (63 Senior Engineer vs 62 GPT-5.5, jumping from 33.5 on Opus 4.7; 79.6 writing, a new high with the fewest AI tells of any non-Claude model) is useful external corroboration when reasoning about which model to route RDCO work through — both agent/coding tasks and Sanity Check prose. Their "model outpaces the app around it" caveat reinforces RDCO's own bet: the leverage is in the harness/scaffolding around the model (skills, context management, sub-agent fan-out), not the raw model app.

• Compound engineering v2 (4→7 steps; 43 subagents, 38 slash-skills) → RDCO's skill/sub-agent architecture. Kieran's "ideate and plan to the front, polish to the end, AI in the middle" sandwich is a direct cousin to RDCO's IC-mode-vs-production-mode split and the multi-seat pipeline (spec → test → code → critic). The "humans as bread" framing maps to the founder↔COO operating rhythm: founder sets intent and judges output, Ray runs the middle. Worth diffing their 7-step loop against RDCO's 12-stage production workflow.

• Codex-for-knowledge-work guide (context files, rules, review checklists, five levels) → harness/context-management thesis. The "context files + rules + review checklists make agents reliable across a workday" claim is the same load-bearing idea as RDCO's CLAUDE.md hard rules, working-context.md scratchpad, and the context-rot discipline of routing long artifacts through sub-agents. Their "five levels of use, one-off → compounding systems" ladder rhymes with the L4→L5 north-star framing.

• "After Automation" / "After After Automation" → already-filed RDCO thesis material. Both essays are already in the vault; this roundup is the curated entry point that ties them together with the podcast. The "more work, not less" thesis is congruent with RDCO building toward an unhobbled COO agent rather than treating automation as headcount reduction.

• Doctronic AI-prescribing pilot → human-gated review pattern + founder's health context. The "recommend-or-escalate, human signs off" architecture is the same shape as RDCO's no-autonomous-external-email and PR-self-merge-with-gates patterns: the AI proposes, escalates uncertainty, a human ratifies the risky calls. The "nervous junior, overcaution is good at this stage" framing is a clean argument for conservative defaults in reversible-vs-irreversible action gating. (Tangential personal relevance: founder tracks gout / MASLD / retatrutide — AI-assisted prescription-renewal pilots are an adjacent watch item, not actionable here.)

• Every as a business model → content-as-product reference. Every continues to be the canonical media+software company RDCO studies: editorial (Context Window, Vibe Check), software (Spiral/Sparkle/Cora/Monologue/Proof), education (camps), and consulting all cross-funnel in a single Sunday issue. This roundup is a clean specimen of how they bundle owned-product cross-promo into editorial without a paid sponsor — relevant to how Sanity Check could eventually monetize without breaking editorial trust.

Related

• [[2026-02-18-every-vibe-check-sonnet-4-6]] — prior Every Vibe Check (Sonnet 4.6); same model-eval format, useful for benchmark-trend comparison
• [[2026-05-29-every-compound-engineering-upgrade]] — the standalone note on Kieran's compound-engineering v2 that this issue curates
• [[2026-05-21-every-after-automation]] — Dan Shipper's "After Automation" essay referenced here
• [[2026-05-27-every-after-after-automation]] — Katie Parrott's encyclical companion piece curated in this issue
• [[2026-02-23-every-chatgpt-memory-context-rot]] — context-rot / context-management thread that the Codex-knowledge-work guide extends
• [[synthesis-harness-thesis-dissent-2026-04-12]] — RDCO's harness-vs-model synthesis, which the "model outpaces the app" framing reinforces