06-reference

every codex of ones own

2026-08-06·reference·source: Every·by Katie Parrott
codexai-agentspersonalizationagent-configurationagentic-workflowknowledge-work

"A Codex of One's Own" — Every / Katie Parrott

Why this is in the vault

Every's Context Window issue argues that a good Codex (agentic AI) setup can't be copied wholesale — it's built around a specific person's workload, thinking style, and life, not a generic job description — using two Every staffers' divergent Codex configs as the case study. The full email body rendered cleanly, but the "Inside Every" piece is itself a free preview of a subscribers-only post — the workspace-design interview prompt it promises ("If I were starting over, I'd use this prompt...") sits behind Every's paywall and was not retrieved.

Mapping against Ray Data Co

Direct validation of a decision RDCO has already made and logged in memory: the [[project_personal_license_boundary|personal-license boundary]] keeps Ray's CLAUDE.md, SOPs, and MEMORY.md deliberately founder-specific rather than genericized into a resellable "AI COO template." Every's thesis — "you can borrow someone's ideas but not their workload, their brain, or their life" — is the same argument against templating an agent config for someone else's operation. Where it adds something new: Every's example (two staffers with the same tool converging on opposite setups — one heavy on planning/context/supervision, the other on fire-and-forget reminders) is a useful frame for future hires or collaborators on Ray — the instinct to copy Ray's exact CLAUDE.md wholesale onto a second person's workflow would be the mistake this piece warns against, even inside the same company.

The Signal section (DeepMind CEO Demis Hassabis moving to Alphabet chief scientist, Koray Kavukcuoglu taking the CEO seat; Meta launching Muse Code, a terminal coding agent) is dated competitive-landscape tracking, not load-bearing for RDCO strategy — noted for completeness, not analyzed further here.

The core argument

Codex setups are idiosyncratic because they encode how a person thinks and works, not just what tools they use. Every's head of ops and head of consulting compared notes and found opposite approaches — one optimized for "minimal process while still executing," the other for detailed planning and supervised output — despite using the same underlying tool. The piece's own resolution: instead of copying someone else's setup, have the agent interview you about your work and decisions, then let it propose the workspace architecture. The actual interview prompt is paywalled.

Curation section

"Links worth a click" — three research items, none deep-fetched (see below):

No deep-fetches taken. The newsletter's own blurbs already carry the load-bearing fact for each item (the MerchantBench 27.3% figure in particular is the whole hook), so a third-party fetch would add color, not mapping strength — RDCO relevance here is real but secondary to the main piece, and the cap-2 budget was better saved for a stronger hit.

⚠️ Sponsorship

This issue carries an explicit paid sponsor block for Attio ("the CRM for teams who set the pace"), plus Every's routine footer self-promo of its own AI product bundle (Sparkle, Cora, Spiral, Monologue) and a subscription upsell gating the full "Inside Every" piece. The Attio placement is a standard newsletter ad slot, unrelated to the article's argument — no evidence it shaped the editorial content. The footer bundle promo is Every marketing its own tools; treat both as normal newsletter monetization, not bias in the Codex-personalization argument itself.

Related