Generate TL;DR for articles and long emails.
Every row saved to the article database gets a TLDR posted as a comment on the row — short and bullet-shaped for articles and long emails, long and partner-grade for YouTube podcasts, with the model and rubric chosen automatically based on the URL.
Prompts used
The exact LLM prompts this automation calls — system instructions and user messages, in execution order, with the model each step uses.
- Article / Long Email TLDR (2-sentence + bullets) →1 prompts
Turns any scraped article or long-form email into a tight, scannable TLDR — a 2-sentence summary followed by 3-5 supporting bullets, capped at 1,750 characters so it fits cleanly into a Notion comment, a Slack post, or a CRM note field. Pairs with the Article and Email TLDR automation, which scrapes the URL, branches on content type, and posts the model's output back as a comment on the source row. The prompt assumes the title is already captured upstream, so it deliberately skips re-stating it and goes straight to the substance.
Reference build
A teammate saves something to the article database — a long email, a Substack post, a 60-minute podcast on YouTube. A few seconds later (or, for YouTube, a few minutes after the transcript catches up), a comment appears on that row with the actual signal: a tight TLDR for prose, a partner-grade podcast brief for video. The DB row stays the source; the comment is what people read first.
Runs off the same article database as the SharePoint archive and CRM capture workflows — the database row in Module 5 of the Thematic Work playbook is the single payload all three keys off of.
Podcast brief rubric
What the Claude Opus system prompt actually enforces. The rubric — not the model — is the reason the brief reads like something a partner would write.
| Section | What the model must do |
|---|---|
| HEADLINE TAKEAWAY | 1–2 sentences. Not 'the guest discussed X' — a claim or worldview shift: 'X is happening because Y, and the implication is Z.' |
| KEY INSIGHTS | 3–8 bullets scaled to substance. Each must carry a number, mechanism, contrarian view, named company / technology, or a forecast with reasoning. If a bullet could appear in a summary of any other podcast on the topic, cut it. |
| CONTRARIAN OR SURPRISING CLAIMS | Only if present. Things a sharp investor would want to interrogate. Skip the section if there was nothing surprising — don't manufacture it. |
| WHAT OUR FIRM SHOULD DO WITH THIS | 1–2 sentences, only if warranted. A specific implication: a sector to revisit, a company to look at, a thesis to update, a question to ask a portfolio CEO. Better to omit than to invent. |
Hard rules in the same prompt: scale length to substance, paraphrase by default, capture the reasoning behind a forecast (not just the forecast), don't summarize the host's questions, don't describe the structure of the episode, no meta-recap at the end, no manufactured insight on a thin episode — call it thin and stop.
Gotchas
- 01One trigger, two summary shapes. The Zap fans out on whether the saved URL contains 'youtube'. Articles and long emails get a short, scannable TLDR; YouTube saves get the long-form podcast brief — because a podcast is a 60-minute conversation and a 3-bullet TLDR throws away the value of having read it.
- 02Path B waits 5 minutes on purpose. The YouTube transcript is written back into the same Notion row by a separate workflow (the 'archive curated articles' SharePoint job, which pulls the transcript via Apify). The 5-minute delay is the cheapest way to let that finish before the summarizer reads the row. If transcripts are slow, the delay is the knob to turn — not the prompt.
- 03The model rubric is the product, not the model. The Claude Opus system prompt is opinionated: lead with the insight, paraphrase by default, capture the reasoning behind a forecast (not just the forecast), don't summarize the host's questions, don't write a meta-recap. Most podcast summaries fail because they describe the episode instead of extracting the signal. The rubric is the difference.
- 04Article TLDR is hard-capped at 1,750 characters. The Notion comment is meant to live as an at-a-glance preview, not as the article itself. The prompt enforces 'plain text, 2-sentence summary then 3–5 bullets, no title'. If summaries start running long, the cap moves — not the bullet count.
- 05Comments, not properties. The TLDR lands as a Notion comment on the row, not as a property value. Comments don't blow up the database view, they thread under the row, and they're trivially distinguishable from the body. Trade-off: you can't filter the DB by 'has TLDR' without an extra column.
- 06Title is the join key on path B. The 5-minute delay step looks the row back up by Title against the same database. If the title got renamed in those 5 minutes (a teammate cleaning up the row), the lookup misses. Rare in practice; worth knowing before debugging a 'why is the TLDR missing' report.
- 07Different model per path is intentional. GPT-5 with reasoning=medium is the cheaper, faster path for short article work; Claude Opus is the spend-more, get-more path for podcasts where the rubric leans on long-context synthesis. Either model can do either job — the split is about per-call economics, not capability.
- 08Pairs with Notion AI for chat. Once the transcript lives in the row and the brief lives in a comment, anyone reading it in Notion can ask Notion AI follow-up questions against the full transcript without leaving the page. The TLDR is the entry point; the transcript + Notion AI is the deep dive.