The content agent (ChatGPT) authors packs from these guides, and the repo has
documented LLM over-compliance failures: snack overload, 416/499 same-mechanic
six-question blocks, exactly-6-of-each mechanic counts. The guides banned
therapy voice but never the machine-voice failure modes.
- QUESTION_CONTENT_GUIDE.md: new section 8A "Write Like a Human" — AI-tell
phrase ban, no-exclamation/no-emoji/sentence-case rules, humor craft (joke
lives in the detail not the adjective; one absurd detail max; adult-deadpan
mock-formality register; punch at the situation never the partner; snort
test at ~1-in-4), anti-uniformity guards (guide examples are categories not
vocabulary; no template blocks in any 10 consecutive; uneven distribution
is correct), texting test. Also retitled section 18 agent-neutral.
- DAILY_SINGLE_CHOICE_WEEKDAY_SYSTEM.md (v8→v9): "Daily Voice: Human and
Funny" subsection — weekday theme is the angle not a template stem;
wildcard absurdity rules; points at 8A as authority.
- QUESTION_QUALITY_CHECKLIST.md (v10→v11): machine-voice Automatic Rejects,
implausibly-even mechanic-count repetition check, and 5 new marking reasons
(ai_voice, labeled_joke, uniform_option_grammar, guide_example_parrot,
brand_caption_voice) so patch discipline can target these failures.
Additive only: no schema/importer-contract, count, or patch-policy changes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
fb9a876b98