AI hooks can promise more than your post delivers.
What the Reel says
AI hooks can promise more than your post delivers.
AI writes your post's first line: the hook.
10
drafts for bakery slides
One offered “5 days of better sourdough.”
The slides covered storage for 3 days.
GPT-5.5 graded 10 new ones itself: all Yes.
A separate prompt caught that promise.
Check in a new chat. Read every Yes.
Sources and assumptions
- Our test: GPT-5.5 (OpenAI API), run 'before' (record.tests): GPT-5.5's real 10 hooks for the fictional Rise Street Bakery sourdough carousel, including hook 5 '5 days of better sourdough starts with this storage trick.' The slides never mention 5 days. (checked 2026-11-05)
- Our test: GPT-5.5 (OpenAI API), run 'after' (record.tests): When the same prompt asked for hooks, each hook's promise and whether the slides keep it, GPT-5.5 wrote a new list and marked all 10 'Yes'. That included '**The 2-second trick for reviving stale sourdough**' and '**Your sourdough is going stale too fast—here's why**'. (checked 2026-11-05)
- Our test: GPT-5.5 (OpenAI API), run 'strict-check' (record.tests): A separate prompt that checked the first list against the slides marked hook 5 'No. Missing a clear “5 days” structure or day-by-day guidance through Day 5.' It marked the other 9 Yes, including 'stay crusty for days'. (checked 2026-11-05)
- Anthropic: Prompting best practices (Claude docs): 'Ask Claude to self-check. Append something like “Before you finish, verify your answer against [test criteria].”' This is general advice to check output against criteria. It is Anthropic's guidance for Claude, not OpenAI's. (checked 2026-11-05)
Assumptions:
- The bakery (Rise Street Bakery) and its 5-slide sourdough carousel are fictional sample material we wrote. Its storage tips are the bakery's draft copy, not facts the post teaches.
- Each test is one GPT-5.5 run through the OpenAI API, not the ChatGPT app. Label it 'GPT-5.5 (OpenAI API), Nov 2026'. A rerun will give different hooks.
- The tests ran on the evening of 2026-11-05 Pacific (ran_at shows 2026-11-06 UTC).
- Saying that 'The 2-second trick' and 'here's why' overstate the slides is our own reading, not a model output. On screen it must read as our judgment, with the slide text shown: slide 5 says 2 seconds under the tap, then 5 minutes at 350°F; no slide explains why bread goes stale.
- The queue idea expected three undeliverable hooks. The real runs gave one clear broken promise (hook 5, '5 days') that the separate check flagged, plus two stretches that GPT-5.5 marked Yes. The post must report exactly that and never say 'three'.
- The OpenAI prompt-engineering guide in the queue's check note could not back a 'check output against criteria' claim. The GPT-5.5 prompting-guide URL now serves a GPT-6 page. Neither is cited. Anthropic's best-practices page is the only source used for the self-check idea, and it is attributed to Anthropic.


