Lesson 69.3 got people talking. Now protect the conversation. Communities rarely die from one dramatic incident; they fade when low-effort posts, subtle spam, and unresolved friction teach good members that posting is not worth it. The fix is norms written early, enforced consistently, with AI doing the sorting and humans owning every removal.

Norms are behavior, not vibes

"Be kind" is a wish, not a norm. A norm names observable behavior and the consequence:

• Show your work: questions include context, what you tried, and the source or photo. Bare "how do I…?" posts get one redirect with a template, then removal.
• Critique the work, not the person: "this paragraph buries the source" is fine; "you are sloppy" is not. First offense: edit plus warning. Second: timeout.
• No drive-by promotion: links require a substantive write-up of what you learned. Link-only posts are removed with an invitation to repost properly.
• Stay on-job: the purpose statement's "not-this" list is enforced. Off-topic threads are moved or closed with a pointer, not debated.

Keep it to five to seven norms, each with an example of good and bad. Post them where posting happens, not in a forgotten legal page. Every norm should answer "what do I do instead?" — norms that only forbid, without modeling, breed resentment.

Write norms in your own voice, briefly. A tailor's circle might say: "Post the deadline, the fabric, and a photo. 'Can someone fix this?' with no photo gets one reminder, then removal." Specificity is kindness — newcomers succeed faster when the bar is visible.

AI sorts; humans decide

AI is an excellent triage clerk and a terrible judge. Use it for what it does well:

• Sort: classify incoming posts (question, showcase, off-topic, possible spam) and route them.
• Summarize: condense long threads into "question, attempts, current best answer" for moderators.
• Flag: surface likely spam, duplicate questions, or heated language for human review — with quotes, never verdicts.
• Draft: propose a moderation reply in the community's voice for a human to edit and send.

Humans own three things absolutely: removals, bans, and rule changes. Every removal gets a human-written reason citing the specific norm, plus a path back ("repost with the source passage included"). Log removals in a simple table — date, thread, norm cited, action — and review monthly. Patterns in the log ("half our removals are link-only posts") mean the norm or the onboarding needs work, not that members are bad.

Never auto-delete or auto-ban based on an AI score, and never present an AI summary as the reason for action. Quote the actual post. Members accept "removed under Norm 3: links need a write-up" far better than "flagged by our system." The standard is procedural fairness a smart newcomer could explain back.

The moderation decision tree

Teach every moderator — even if that is just you — one tree:

New post arrives
 ├─ Spam / scam? → remove + ban, log it
 ├─ On-job and meets format? → leave, maybe highlight
 ├─ On-job but low-effort? → reply with template + 48h to fix, then remove
 ├─ Off-topic but valuable? → move to right place or close with pointer
 └─ Heated / personal? → pause thread, message parties privately, edit or remove the personal part

Two principles sit behind the tree. Escalate gradually: warn before removing, remove before banning, except for scams and abuse where speed protects members. Moderate in public, discipline in private: the norm reminder can be visible (it teaches lurkers), but arguments about a specific member never are. Take it to messages, decide, then post a one-line outcome if the thread needs closure.

Plan for the first hard case now: a popular early member who brings energy but ignores the format. Decide in advance that norms outrank popularity. Granting exceptions to favorites is the fastest way to teach everyone else that norms are theater.

Exercise: publish norms and the tree

Create NORMS.md with your five-to-seven norms, each with a good/bad example and a first/second consequence. Then add a MODERATION-TREE.md: the tree above adapted to your community, naming who moderates, maximum response time, and where the removal log lives. Draft three AI-triage prompts: a classifier ("label this post: question/showcase/off-topic/spam, quote the deciding line"), a summarizer ("summarize this thread as question/attempts/best-answer/open"), and a flagger ("list norm numbers possibly violated, quote evidence, recommend human review — do not decide").

Finish line: posted norms, a named moderator, and a triage workflow where no AI output removes anything without human approval.

Verify quickly: hand a borderline post to a friend with only your norms and tree. If they reach the same decision you would, the docs work.

Common failure mode: ten vague rules enforced inconsistently. Fewer, sharper norms enforced every time beat a constitution nobody reads.

Check your understanding

1. What makes a norm enforceable rather than decorative? 2. Which moderation tasks can AI own, and which must humans own? 3. Why should discipline happen privately while norm reminders stay visible?

Next

Class 69 built the room and its rules. Class 70 turns conversation into capital — designing contributions members want to make, organizing them with AI, and converting the best into searchable assets.