Why this is in the vault
Every's "Context Window" issue profiles a practitioner (Every's head of tech consulting, Mike Taylor) who has Claude ghostwrite most of his posts and dictates/edits rather than drafts — plus a wrap-up of the Anthropic watermarking backlash and a Codex batch-editing workflow — worth keeping as a practitioner-level counterpoint to the "writing is thinking" objection to AI-assisted writing.
Mapping against Ray Data Co
The lead item bears directly on how Sanity Check should be produced: Mike Taylor's stance — "if I've run the test, it's completely valid for me to go to an AI and say, 'I ran a test, here's what I did'" — is functionally the same authorship claim Ray Data Co already makes for Sanity Check drafts (founder-run analysis, AI-assisted prose, human review gate before publish). It's evidence the "outsourcing drafting = outsourcing judgment" critique doesn't hold when the underlying reasoning (the test, the data, the position) is human-original and AI only handles formatting/expression — directly relevant to the /draft-review and IC-mode-vs-production-mode SOPs, and a useful citable counterpoint if the "was this written by AI" question ever comes up about RDCO content. Secondarily, the Signal section on the Anthropic watermarking backlash adds a third source (after Stratechery and Ship30for30, both already filed) converging on the same take: the watermark can't distinguish "AI ran the keyboard" from "AI helped format a human's own findings" — which is precisely the gap Sanity Check's production model sits in, and reinforces that watermark detection is not something RDCO needs to defensively design around.
The core argument
Mike Taylor (Every's head of tech consulting) has Claude ghostwrite most of his posts, including his most popular Every piece; he revises structure but doesn't draft the bulk of the text himself. His line: writing isn't the only way to think — he prioritizes getting real, tested findings in front of people quickly (a full day of writing collapses to two hours) over prose craft, and refuses to let AI write about anything he hasn't personally done or tested. The issue's Signal section recaps the Anthropic SynthID-based watermarking rollout and backlash: watermarking doesn't reduce output quality (it preserves token-probability odds on average) but also can't distinguish heavy human editing from raw AI output, can be defeated by rewriting or routing through another model, and risks stigmatizing AI-assisted work regardless of how much human judgment went into it.
⚠️ Sponsorship
This issue carries an explicit sponsor block for Brief (briefhq.ai), a product-prototyping/spec tool, with a "Want to sponsor Every?" house CTA immediately after it. The sponsor content is siloed to its own block and doesn't touch the watermarking or AI-writing sections used above, but flagging per policy: treat the "customers can't use prototypes" framing as vendor-supplied, not Every's independent editorial judgment.
Related
- [[2026-08-12-stratechery-anthropic-watermarking]]
- [[2026-08-19-ship30for30-watermarking-implications]]
- [[2026-04-06-every-writing-with-ai-harder]]