Summarize a YouTuber's latest video into a Google Doc daily
A Make scenario runs at 9 AM, finds a YouTuber's newest video, has Hugging Face summarize it, and appends the summary to a Google Doc.
Cursor Agent writes and maintains interlinked markdown pages from user-supplied sources, summarizing and cross-referencing, with Obsidian used to browse.
Cursor Agent writes and maintains interlinked markdown pages from user-supplied sources, summarizing and cross-referencing, with Obsidian used to browse.
The vault folder is opened in both Cursor and Obsidian. Sources go into an immutable raw/ folder, and an AGENTS.md file sets the wiki's conventions. The user tells the agent to ingest one source. It writes a summary, updates the index and related pages, and appends to a log. The user reads the summaries and steers. Later, lint passes check for contradictions and orphan pages.
The source is a pattern description, not a report with results. It says one source can touch 10–15 pages and that the index approach works at roughly 100 sources without search infrastructure.
Ingest sources one at a time and review the summaries. The author does this, though batch ingest with less supervision is possible. Add a search tool such as qmd once the index stops being enough.
Read the sources in [folder of papers and articles] on [topic] and build a wiki of interlinked markdown pages in [wiki folder]: one page per [concept / paper / person], each with a summary and links to related pages. Add an index page. Finish when every source is covered and links resolve, and ask me before overwriting any existing file.
A Make scenario runs at 9 AM, finds a YouTuber's newest video, has Hugging Face summarize it, and appends the summary to a Google Doc.
An n8n workflow scrapes Reddit posts and comments via Bright Data, extracts concerns and themes with Gemini, and writes reports to Notion.
A user left a Gemini agent running overnight to analyze 170k LLM logs with the llm CLI and a free model, and it finished the job despite limits.