@Kenzie Vick thanks dude!

@Kenzie Vick thanks!

### - HOT TIP! use Google Apps Scripts ------------------------------------------------------------------ -Ask your AI agent if this problem can be solved using Google Apps Scripts/Functions. https://script.google.com/ -If it can, ask them to create you the script and walk you through the steps on how to run it and deploy it. -It gets native Sheets/Gmail/Slides access with NO 'cloud project' or 'billing'.

@Yao Yong The answer is dependent on your scenario and what your AI did since it's nondeterministic. So I can only speak to how 'my project' works. How mine works (no LLM API calls) - youtube_api.py → pulls metadata from the YouTube Data API v3 (titles, tags, views, dates, channel stats). No LLM. - harvest_youtube.py → ranks by views/day velocity and tags each video by keyword matching (topic buckets, tool names, phase) on the title + tags. It also pulls transcripts via youtube-transcript-api (free, no key) — but only to keyword-scan them for tool/technique mentions (e.g. confirm "Civils.ai" or "ODA File Converter" actually came up), not to summarize them. - analyze_youtube.py → its own docstring says "Deterministic, no API calls." It counts, ranks, maps the 10 industry phases, and finds white space. Recommendations come from a hard-coded ROADMAP_SEED dict (one canned line per phase). So nothing in the tools sends video content to Claude/GPT/Perplexity for summarization. It runs free, fast, and repeatable — which is exactly why it's schedulable. Where the "LLM" actually is: The one place intelligence/prose enters is the agent (me/Claude) at deck-build time — the analysis output is "structured numbers + seed recommendations; the agent refines the narrative when building the deck." That's the WAT pattern: the Tools stay deterministic, the Agent supplies the judgment in the loop. There's no LLM baked into the deployed scripts (which is good — it keeps cost/nondeterminism out of the automation). The important nuance: Because of this, the deck shows signal, not synopsis — per video you get title, channel, velocity, tools mentioned, and phase, plus aggregate trends and white-space. It does not produce an abstractive "here's what this video teaches" summary of each video. The transcripts are mined for keywords, not read-and-summarized.

Just started. Lets see what we can build.

I used CLAUDE.md but I didn't see any requirements about the newsletter demo. Did I miss a step? I mean I can guess but...