Lesson 70.1 produced steady contributions in a consistent format. Now volume becomes the problem: twenty teardowns a week is wonderful and unreadable. AI's proper role here is librarian, not author — classify, summarize, tag, and draft digests while originals stay untouched and a human reviews everything before it represents the community.

The division of labor

Draw the line sharply. AI may classify (this thread is a filing teardown, that one is tooling chat), summarize (condense a 40-reply thread to question, attempts, best answer), tag (filing type, difficulty, method), detect trends (five separate threads hit the same footnote problem), and draft (a digest or newsletter version for review). AI may never rewrite a member's words in place, attribute a claim a member did not make, or publish under a member's name.

Three preservation rules follow. Originals are immutable: AI-generated summaries sit alongside the thread, clearly labeled, linked — never replacing the member's text. Attribution travels: every summarized point carries the author's name and a link to the exact message. Corrections flow upward: if a summary errs, fix the summary and note the fix; never silently edit the original to match the summary. Members lend you their words. Librarians do not rewrite the books.

This matters practically: the outline lists "turn discussions into articles" and "create newsletters" as AI powers. Both are real, and both become exploitation or misinformation the moment a draft ships without review. The workflow below makes review the bottleneck on purpose.

The weekly digest workflow

Run this every week, same day, same shape:

Step 1 — Collect. Gather the week's contributions (links, not copies). An AI classifier labels each: keep (meets the card format), promising (good core, needs a nudge), or chat (valuable socially, not archivable). Humans spot-check the labels — ten minutes, not two hours.

Step 2 — Condense. For each keep, AI drafts a three-line abstract: problem, approach, link. For long threads, a slightly longer brief: question, attempts, current best answer, open issue. Each abstract ends with author credit and a deep link. Nothing is paraphrased beyond recognition; quotes stay quoted.

Step 3 — Detect. Ask AI for a trends note: which tags spiked, which questions recurred, which answers conflict. Treat this as a hypothesis list, not findings. A human confirms ("yes, three threads really did hit revenue-recognition footnotes") before it goes further.

Step 4 — Draft. AI assembles a digest draft: three best contributions with abstracts, one trend worth watching, one open question for next week. The draft is explicitly marked unreviewed.

Step 5 — Human review. A named reviewer checks every abstract against its original, verifies attribution, cuts anything the author would not endorse, and signs the digest. Only then does it publish — to the archive, the newsletter, or both. Log the review: date, reviewer, items included, items cut.

Total human time: roughly an hour a week for a mid-size community. That hour is the trust budget. Spend it.

Prompts that behave

Three reusable prompts, each with guardrails baked in:

• Classifier: "Label each thread QUESTION / SHOWCASE / CHAT / SPAM. Quote the single line that decides the label. Confidence low/medium/high. Do not rewrite anything."
• Summarizer: "Summarize this thread as: Question (one line) / Attempts (bullets with author names) / Best answer so far (quote + author + link) / Still open (one line). Preserve original wording in quotes. Flag any claim you cannot trace to a message."
• Trend watcher: "List tags appearing 3+ times this week with thread links. For each, write one cautious sentence starting 'Members discussed…' Do not generalize beyond the linked threads."

Notice what each prompt demands: quotes, links, hedged language. AI output without evidence pointers is a liability, not a deliverable.

Handle disagreement explicitly, because the most valuable threads contain it. When two members offer conflicting approaches — different filing interpretations, different tooling recommendations — the summary must preserve both positions with their reasons and supporters, not declare a winner the community did not agree on. Write the digest entry as "A argues X because…; B argues Y because…; unresolved: Z." Forcing false consensus to make a tidy digest teaches experts that nuance gets flattened here, and they stop posting it. A digest that faithfully records open questions is more trustworthy than one that pretends every thread resolved neatly on Friday.

Exercise: ship one reviewed digest

Run the workflow once, end to end, and save it as DIGEST-001.md: dated, with abstracts, author credits, links, a trends note marked confirmed-or-not, and a review sign-off line ("Reviewed by [name], [date]; [n] items cut"). Include a short DIGEST-WORKFLOW.md beside it: your five steps, prompt texts, and time log.

Finish line: one published digest a member could forward proudly, with every claim traceable to an original they can open.

Verify quickly: pick any abstract and click through to the original. If the meaning shifted or the credit is wrong, fix the workflow before scaling.

Common failure mode: auto-publishing unreviewed summaries. The week you skip review to save an hour is the week a misattributed claim teaches members to stop posting.

Check your understanding

1. Which five organizing tasks can AI own, and what must it never do? 2. What are the three preservation rules for member originals? 3. What does the human reviewer check before a digest ships?

Next

Weekly digests keep the present readable. Lesson 70.3 makes the best of it permanent — a publish-or-not gate that turns selected discussions into searchable assets without exploiting contributors.