The software research kept turning up the right pattern in the wrong domain — progressive disclosure applied inward to developer docs and agent tooling, never outward to public writing. The actual unlock came from a different direction: Zack’s own legal background, and the observation that CRAC/CREAC already do exactly this. The conclusion, which bookends a CRAC/CREAC paragraph, is level 1 — the answer, for a reader who’s busy and may stop there. The rule (and, in CREAC, its explanation) plus the application is level 2 — the reasoning trail, for a reader who wants to check the work. The citation trail is level 3, and it’s the strongest part of the analogy: a verification layer built for a machine process, not narrative reading, mechanized since 1873 via Shepard’s Citations. That’s a non-narrative reader served by legal writing for a century and a half before “AI agent” was a phrase anyone used.

That parallel is what turned an interesting reframe into an actual thesis: legal writing already solved this, and the rest of communication is only now catching up. It also settled the piece’s structure — open cold on the historical fact, unpack the three layers, turn the observation outward, name the real risk of a misreading agent that reports back with false confidence, dispatch the AEO/llms.txt objection in one paragraph, and demonstrate the site’s own version of the structure live rather than re-describing it. One thing came out of the outline pass rather than into it: progressive disclosure, which had been carrying the argument as recently as the previous entry, got demoted to a one-line aside once the older, better parallel was confirmed. The platform split — full structure on the site, one unexplained line on LinkedIn — closed out the chapter, with one question left open on purpose: whether the LinkedIn post teases the site piece or stands alone.

Main takeaways

  • The stronger analogy didn’t come from adjacent research — it came from the author’s own field, checked rather than assumed.
  • A parallel that predates the technology it’s being compared to (1873, versus a phrase invented in the last few years) is harder to dismiss as trend-chasing.
  • Structure decisions (what leads, what’s demoted to an aside, how long the objection-cutoff gets) are argument decisions, not formatting ones.

What I got wrong

  • Progressive disclosure was drafted as the lead hook before the CREAC parallel surfaced, and that framing had to be cut once the stronger version proved itself — not repurposed, cut.
  • The LinkedIn tease-versus-standalone question was left open deliberately rather than resolved on a guess, and it’s still open as of the last entry in this chapter.