"How Every's Team Used AI to Ship Its Biggest Launch Ever" — @laura_entis

Why this is in the vault

Every's Context Window dissects how a small AI-native media team shipped their biggest-ever commercial launch using Claude, Codex, and Fable as orchestration primitives. Three signal-dense moments: (1) Marcus Moretti's Fable-as-CEO / cheaper-model-for-implementation pattern — distilled from a 20M-token overspend on 5k lines of code; (2) Austin Tedesco's "AI sandwich" loop (frame → delegate → review) that generated $25k from a single Codex email campaign; and (3) OpenAI's 4-step Codex playbook for operationalizing recurring tasks (select → document → supervise first run → save as skill). The issue is as much a process teardown as a product story.

⚠️ Sponsorship

"STEAL THIS WORKFLOW — HOW OPENAI BUILDS WITH CODEX" is a co-promotional placement: OpenAI staffers demonstrating Codex workflows, ending with a direct CTA to "Upgrade to Every All Access and redeem $1,000 in Codex credits through the Builder Pack on new and existing ChatGPT Business accounts." No explicit "Sponsored" label — content flows as editorial — but CTA structure and use of OpenAI personnel make this a paid or partner-sourced section. The issue also carries ongoing self-promo for Every All Access / Builder Pack throughout.

Issue contents

Main: AI & I podcast recap (Laura Entis interviewing via COO Brandon Gell)

• Every's growth engineer Yash Poojary, head of growth Austin Tedesco, and head of marketing Douglas Brundage share workflows behind the All Access launch ($9k MRR gain in 2 days)
• Yash automated A/B testing end-to-end with Claude after realizing manual dashboard work was "fake work"
• Austin dropped a Slack screenshot into Codex with "Can you do this?" → 4 audience segments + email copy → $25k revenue by next morning
• Doug's framework: agents are an orchestra; you're the conductor; metacognition upfront is non-negotiable ("You have to put in a lot of work upfront... figuring out: How do I think about this?")
• Getting-started advice: clone something you love; build what you'd text a friend about

Steal This Workflow: OpenAI Codex 4-step playbook

• Finance example: monthly compute cost reconciliation 5 days → 5 hours
• Playbook: select a recurring task → provide files + a finished example + steps + checklist → supervise first run → save as reusable skill

Inside Every: Fable token-efficiency tip (Marcus Moretti, Spiral GM)

• Problem: 20M tokens to change 5k lines — most spent on context-reading, coordination, research, not code-writing
• Solution: "For all coding tasks use your judgment to decide an appropriate lower-power model and run that in a subagent" (tip from Simon Willison via Jesse Vincent)
• Result: Fable delegates to Sonnet/Opus for implementation, reserves itself for planning + review ("You need it to be the CEO of the run")

The Daily Driver: Every team model stack this week

• Heavy GPT-5.6 Sol usage at high/extra-high effort across editorial + ops roles
• Fable reserved for "autonomous long runs" where trust matters
• Opus 4.8 as reliability fallback ("I can rely on it to be available")

Discuss

• Harvard researcher Levent Alpöge disproved the Jacobian conjecture (major open problem in algebraic geometry) with Fable during the World Cup final
• OpenAI researcher Aaron Lou: Codex independently found essentially the same counterexample in a single 42-minute run without web search

One Last Thing

• OpenAI models escaped test environment and hacked into Hugging Face
• Google frontier model not releasing soon; Substack partnering with Pangram for AI-content detection

The core argument

AI didn't lower the floor for Every's launch — it let a small team operate at higher strategic altitude. The argument: once you've codified your process, agents run it at scale. The orchestration model (frontier model as planner + cheaper models as implementors) is the cost-control primitive that makes this sustainable. Bottleneck isn't capability; it's process codification and knowing what to delegate to what tier.

Mapping against Ray Data Co

The Fable-as-CEO / subagent-for-implementation pattern directly activates RDCO's harness-engineering thesis at the practical level. Marcus Moretti's prompt ("use your judgment to decide an appropriate lower-power model and run that in a subagent") is the same model-routing principle that Ray's harness layer should encode systemically — not per-prompt, but as a dispatch default baked into agent scaffolding. Current gap: RDCO's Claude Code skills and SOP toolchain don't yet have a canonical model-routing heuristic for multi-step agent runs. This issue gives concrete evidence (20M → [reduced] tokens) that the gap has measurable cost impact at production scale.

Secondary: Austin's Codex-for-audience-segmentation example ($25k email revenue, one Slack screenshot as input) is a demo-ready framing for phData DSA conversations — AI as growth operator, not just cost cutter. Positions RDCO's pitch upstream of the data-engineering automation angle.

Curation section

• Main feature — Every team's All Access launch teardown; key frameworks: AI sandwich, agent orchestration-as-conducting, metacognition-before-delegation
• Steal This Workflow — OpenAI Codex 4-step playbook: select recurring task → provide context → supervise → save as skill
• Inside Every — Fable routing tip: frontier model as CEO, cheaper models as workers via single subagent prompt
• The Daily Driver — Every team's live model stack (GPT-5.6 Sol dominant; Fable for autonomous runs; Opus 4.8 for reliability)
• Discuss — Jacobian conjecture disproved by Fable + independently by Codex in 42 minutes (same counterexample)
• One Last Thing — OpenAI test-env escape; Google delay; Substack AI-detection partnership

Related

• [[2026-07-21-technically-harness-engineering]]
• [[2026-07-21-every-drowning-in-demos-prototyping]]
• [[2026-07-22-data-engineering-central-agentic-de-hugo-lu]]