"Opus 4.7 Reels Us Back In" — @Laura Entis

Issue contents

Every's Context Window curation piece argues Opus 4.7 has won back Every's team after a GPT-5.5/Codex-heavy stretch. Headline framing: Codex feels fast but thin; Opus 4.7 functions like a senior magazine editor — better planning, better creative depth — and the new 2.5x fast mode closes the speed gap that pushed people to Codex in the first place. Head of growth Austin Tedesco's test on a creative writing project landed the comparison: Codex behaves like an AP fact-checker; Opus 4.7 plans, structures, and revises. Engineer Paridhi Agarwal flagged that Opus 4.7 proactively suggested parallelizing her work across multiple terminals — an unprompted operational improvement.

Three other items in the issue:

1. Mini Shai-Hulud breach. Attackers compromised 42 official TanStack npm packages via malicious pull requests that hijacked the build system itself (not credential theft directly). The injected code hunted for cloud keys, GitHub tokens, and npm access tokens, with a dead-man's switch that could wipe developer home directories if tokens were revoked mid-attack. Same tactic then hit UiPath and Mistral AI packages. Takeaway: ready-to-run automated repo audits the moment a supply-chain breach is disclosed.

2. AI-isms elimination. Spiral (Every Studio writing tool) added a Gemini 2.5 Flash "top edit" pass that strips telltale AI-generated markers — em dashes, "It's not X, it's Y" reframes, overused vocabulary like "shift," "shape," "delve" — and cut user complaints about AI-feel by 30%. Maintains a crowdsourced list of tells.

3. Defining "agent". Piece argues the useful split is collaborative agents (extend user capability, you stay in the loop) vs delegated agents (execute independently, you don't interfere). Model selection should follow that axis — not raw benchmark scores.

Why this is in the vault

Opus 4.7 is the model I'm running on right now. Third-party reads on its relative strengths matter for two concrete RDCO decisions: (a) which model to default to for which agent surface, and (b) how to frame the L5 capability buildout — the collaborative-vs-delegated split maps cleanly onto where I need more autonomy (delegated mode) vs where the founder wants tight loops (collaborative mode). The AI-isms note also reinforces the founder's standing no-em-dashes rule — Every's data point (30% drop in complaints) gives that rule external evidence beyond personal taste.

Mapping against Ray Data Co

• Model selection by mode, not by benchmark. I currently default to Opus 4.7 across the board. Every's framing suggests an explicit split: keep Opus for planning, drafting, judgment-heavy delegated work; consider Codex/GPT for fast-but-shallow execution tasks where structure is already locked. Worth a deliberate audit of which RDCO skills actually need Opus depth vs could ship cheaper. Not urgent — model cost isn't a current bottleneck — but the framework is portable.
• Collaborative vs delegated agent split. This matches the L5 north star framing in [[project_l5_north_star_strategic_direction]] — the unhobbling is specifically about moving more surfaces from collaborative to delegated. Every's lexicon is useful shorthand for that transition.
• Supply-chain auto-audit. RDCO repos use npm/pnpm packages including TanStack-adjacent ecosystem code. Worth a one-time check that no compromised packages landed in lockfiles, and worth setting up a standing audit hook that runs on breach-disclosure RSS/feed events. Lower priority than the model-selection question.
• AI-isms top-edit pattern. External validation for the no-em-dashes rule ([[feedback_no_em_dashes]]). The broader pattern — a final pass that strips AI tells with a cheap fast model — is portable to Sanity Check drafts and any newsletter copy. Worth a small experiment: run a Gemini 2.5 Flash or Haiku 4.7 pass on the next SC draft after voice-match and see whether it catches anything voice-match missed.

Related

• [[2026-05-03-every-context-window-codex-goes-to-work]]
• [[2026-04-30-every-who-isnt-using-gpt-5-5]]
• [[2026-05-07-every-anthropic-2026-developer-conference]]
• [[2026-05-10-every-ai-work-splitting-in-two]]
• [[feedback_no_em_dashes]]
• [[project_l5_north_star_strategic_direction]]

Decision-worthy

DECISION: Audit RDCO skill roster for which surfaces actually need Opus depth vs could ship on a faster cheaper model (Codex/Haiku/Flash) — and pilot a post-draft "AI-isms strip" pass on the next Sanity Check draft to test Every's 30%-complaint-drop claim against the founder's voice.