docs(seed): add anti-AI-voice + humor-craft standard to question guides

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>

seed/questions/DAILY_SINGLE_CHOICE_WEEKDAY_SYSTEM.md

@@ -1,4 +1,4 @@
-# Daily Single Choice Weekday System v8 — Importer-Aligned
+# Daily Single Choice Weekday System v9 — Importer-Aligned, Human-Voice
 
 This document defines the Closer daily weekday question pack.
 
@@ -304,6 +304,29 @@ Avoid these in daily questions:
 
 These words push the pack toward therapy voice.
 
+## Daily Voice: Human and Funny
+
+The daily pack is the funniest surface in the app, so the "Write Like a Human" standard in
+`QUESTION_CONTENT_GUIDE.md` (section 8A) binds hardest here. Read it before writing any daily batch.
+
+Daily-specific applications:
+
+* The joke must live in the detail, never in a label. If "silly" or "ridiculous" is doing the humor
+  work, the line has no humor. Cut the adjective and make the option specific enough to be funny alone.
+* The house register is adult deadpan: mock-formality and fake stakes for tiny real things. Not
+  chaotic-quirky, not children's-entertainer energy.
+* The weekday theme is the *angle*, not a template stem. Every Thursday prompt opening
+  "What's the funniest..." is a machine fingerprint. Approach each weekday's mood from a different
+  direction every time.
+* Wildcards get the single most absurd detail in the pack — still adult deadpan, still one absurd
+  detail maximum.
+* The guides' example nouns ("two-song kitchen dance", "dessert walk", "snack board") are categories,
+  not vocabulary. Shipping them verbatim is a reject.
+* No exclamation marks in prompts. No emoji in prompts or options. Options in sentence case.
+* Aim for roughly one question in four to pass the snort test (partner would snort and show the
+  other the phone); the rest land warm and concrete. A pack that strains for a laugh on every line
+  is as exhausting as one with none.
+
 ## Good Daily Examples
 
 Good:

seed/questions/QUESTION_CONTENT_GUIDE.md

@@ -564,6 +564,131 @@ What is one small thing that makes you want to be closer to me?
 
 ---
 
+# 8A. Write Like a Human
+
+Section 8 bans therapy voice. This section bans the other failure mode: content that is technically
+warm and playful but reads machine-written. These are the patterns that make a pack feel like an app
+talking instead of a person.
+
+## AI-tell phrases are banned
+
+Never use these in prompts or options:
+
+```text
+spice things up
+take it to the next level
+level up
+unleash
+embark
+elevate
+epic
+ultimate
+game-changer
+adventure awaits
+cozy up (as a command)
+delight / delightful
+whimsical
+a dash of
+sprinkle
+```
+
+These are marketing-bot words. A real person deciding what to do tonight does not say "let's elevate
+our evening."
+
+## Punctuation and case rules
+
+- No exclamation marks in prompts.
+- No emoji in prompts or options.
+- Options use sentence case, never Title Case.
+
+Enthusiasm comes from the idea, not the punctuation.
+
+## The joke is the detail, not the adjective
+
+Never label a joke. "Silly", "hilarious", "ridiculous", and "funny" doing the humor work means there
+is no humor in the line. Make the option itself funny through unexpected specificity.
+
+Dead:
+
+```text
+A silly dance together
+```
+
+Alive:
+
+```text
+A two-song kitchen dance
+```
+
+The second one is funny because "two-song" is a weirdly specific commitment. Nobody labeled it.
+
+If removing the word "silly" or "ridiculous" from a line kills it, the line was already dead.
+
+## One absurd detail, maximum
+
+One unexpected detail makes a line funny. Two makes it a kids' menu.
+
+Good: a formal award ceremony for whoever picked the better snack.
+
+Too much: a formal award ceremony with a trophy made of cheese judged by the cat in a tiny hat.
+
+## The house humor register
+
+Closer's humor is adult deadpan: mock-formality and fake stakes applied to tiny real things.
+
+```text
+official pick
+formal ruling
+lifetime achievement award for couch positioning
+a binding decision
+this week's championship
+```
+
+Not random wackiness. Not chaotic-quirky. Not children's-entertainer energy. Two adults keeping a
+straight face about something small — that is the register.
+
+## Punch at the situation, never at the partner
+
+If the laugh needs a target, the target is the couple's shared circumstance — the fridge, the
+algorithm, the weather, the errand that ate the afternoon. Never either person's body, habits,
+competence, or effort.
+
+## The snort test
+
+The best daily lines make one partner snort and show the other the phone. Aim for roughly one in four
+questions to genuinely hit that bar. The rest should land warm and concrete. A pack where every line
+strains to be funny is as exhausting as a pack with no jokes at all.
+
+## Example lists are categories, not vocabulary
+
+The example lists in these guides (snacks, tiny dates, silly bets...) name *kinds* of content. They
+are not words to reuse. Do not put the guides' own example nouns — "two-song kitchen dance",
+"dessert walk", "snack board", "fake award" as literal text — into shipped questions. Invent your
+own specifics of the same kind.
+
+If a guide example appears verbatim in a pack, that question is a reject.
+
+## No template blocks
+
+Within any 10 consecutive questions:
+
+- no repeated game mechanic
+- no two prompts opening with the same first three words
+- no option set where all options share one grammatical template
+  (four options all shaped "A + adjective + noun" is a machine fingerprint)
+
+## Distribution should look human
+
+A human writer favors some mechanics and neglects others. Exactly N of each mechanic, evenly spaced,
+is a machine fingerprint and a reject signal at review. Uneven is correct.
+
+## The texting test
+
+Read the question as a 6pm text to your partner. If it reads like a brand's Instagram caption or an
+app talking, rewrite it. A person wrote this for one specific other person — it should sound that way.
+
+---
+
 # 9. Tone Rules by Pack Type
 
 ## Fun packs
@@ -1582,9 +1707,11 @@ ECHO Would make a normal couple roll their eyes.
 
 ---
 
-# 18. Claude Instructions for Rewriting JSON Packs
+# 18. Agent Instructions for Rewriting JSON Packs
 
-When Claude rewrites a pack:
+These rules bind every agent that rewrites a pack (ChatGPT, Claude, or any other writer).
+
+When an agent rewrites a pack:
 
 ```bat
 ECHO Read the existing JSON first.
@@ -1607,7 +1734,7 @@ ECHO Report old count, new count, and net catalog change.
 ECHO Report any schema changes.
 ```
 
-Claude should not say a file is done unless:
+The agent should not say a file is done unless:
 
 ```bat
 ECHO The JSON parses.

seed/questions/QUESTION_QUALITY_CHECKLIST.md

@@ -1,4 +1,4 @@
-# Closer Question Quality Checklist v10 — Importer-Aligned
+# Closer Question Quality Checklist v11 — Importer-Aligned, Human-Voice
 
 **See also:** [QUESTION_CONTENT_GUIDE.md](QUESTION_CONTENT_GUIDE.md) | [QUESTION_SCHEMA.md](QUESTION_SCHEMA.md) | [QUESTION_REWRITE_PLAN.md](QUESTION_REWRITE_PLAN.md)
 
@@ -31,6 +31,18 @@ Reject any question that contains or strongly resembles:
 
 These are therapy worksheet patterns.
 
+Also reject any question where (see `QUESTION_CONTENT_GUIDE.md` section 8A):
+
+* an AI-tell phrase appears (`spice things up`, `level up`, `unleash`, `elevate`, `epic`, `ultimate`,
+  `game-changer`, `adventure awaits`, `whimsical`, `a dash of`, `sprinkle`, and the rest of the 8A list)
+* "silly", "hilarious", "ridiculous", or "funny" is doing the humor work instead of a specific detail
+* a prompt contains an exclamation mark, or a prompt/option contains emoji
+* options are in Title Case
+* all options in one set share a single grammatical template
+* a guide example noun ("two-song kitchen dance", "dessert walk", "snack board") ships verbatim
+
+These are machine-voice patterns.
+
 ## Production File Hard Checks
 
 Run these before tone or content review.
@@ -382,6 +394,7 @@ Reject or rewrite if:
 * the same situation repeats with different nouns
 * the same answer pattern repeats
 * the weekday starts to feel like wallpaper
+* mechanic counts are implausibly even (exactly N of each mechanic is a machine fingerprint, not balance)
 
 The pack can pass duplicate checks and still fail repetition review.
 
@@ -433,6 +446,11 @@ Use these reasons when marking weak questions:
 * duplicate_text
 * duplicate_options
 * schema_issue
+* ai_voice
+* labeled_joke
+* uniform_option_grammar
+* guide_example_parrot
+* brand_caption_voice
 
 ## Final Verdict Labels