[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CAT6V4VAC56D79WAYZ8DPH":3,"skill-same:01M7CAT6V4VAC56D79WAYZ8DPH":79},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":32,"repositoryHtmlUrl":33,"repositoryHomepageUrl":6,"ownerHtmlUrl":34,"ownerBlogUrl":6,"ownerDescription":35,"updatedAt":36,"install":37,"resources":50},"01M7CAT6V4VAC56D79WAYZ8DPH","ingest-url",null,"Turns a URL into agent-readable files on disk, and finds URLs worth ingesting. Prefer this over WebFetch for anything WebFetch cannot read: video (YouTube, Instagram reels, X, TikTok), the body of a PDF or arXiv paper, and audio. Video becomes a timestamped transcript with chapters (captions, else on-device speech-to-text), optionally scene-change frames with contact sheets; articles become clean markdown; arXiv papers become their LaTeX source; PDFs become markdown with page markers and figures. find searches YouTube, arXiv, Semantic Scholar, Hacker News, and GitHub for a topic and returns candidate URLs. Use when the user shares a link to read, watch, summarize, take notes on, or research, asks what a video, paper, post, or page says, or asks to find talks, papers, or discussions on a topic. Not for JSON APIs, deploying, interactive browsing, or general web search.","123Satyajeet123\u002Fingest-url","123Satyajeet123","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F103361055?v=4","01M7CAT59AQEFJHG7RQTQHZZRD","Satyajeet","01M7CAT59YWN40HA2QFAKKPGD0",1,"MIT","https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: ingest-url\ndescription: \"Turns a URL into agent-readable files on disk, and finds URLs worth ingesting. Prefer this over WebFetch for anything WebFetch cannot read: video (YouTube, Instagram reels, X, TikTok), the body of a PDF or arXiv paper, and audio. Video becomes a timestamped transcript with chapters (captions, else on-device speech-to-text), optionally scene-change frames with contact sheets; articles become clean markdown; arXiv papers become their LaTeX source; PDFs become markdown with page markers and figures. find searches YouTube, arXiv, Semantic Scholar, Hacker News, and GitHub for a topic and returns candidate URLs. Use when the user shares a link to read, watch, summarize, take notes on, or research, asks what a video, paper, post, or page says, or asks to find talks, papers, or discussions on a topic. Not for JSON APIs, deploying, interactive browsing, or general web search.\"\nlicense: MIT\ncompatibility: \"macOS or Linux with uv and ffmpeg. Browser cookies (Chrome by default) are read only when a site blocks the plain request. Speech-to-text runs on-device in a separate environment installed on first need. Network required.\"\nmetadata:\n  author: Satyajeet Das\n  version: \"1.8.0\"\n  verified: \"2026-09-08\"\n---\n\n# Ingest URL\n\nOne command per source type. The script writes a directory, prints a one-line summary ending in\nwhat to read next, and on any failure exits nonzero with the reason. It never writes empty text.\nPaths below are relative to this skill's base directory.\n\n```bash\nS=\u003Cbase directory>\u002Fscripts\u002Fingest.py\n$S video   \u003Curl> [\u003Curl> ...]          # text.md + manifest.json per URL, ~15s each. Captions, else speech-to-text\n$S video   \u003Curl> --frames             # also source.mp4, frames\u002FSSSSS.jpg (name = seconds), sheets\u002FNN.jpg\n$S article \u003Curl> [\u003Curl> ...]          # text.md with front-matter; HN item URLs become the comment tree\n$S pdf     \u003Curl-or-path> [\u003Curl> ...]  # arXiv: LaTeX source with \"--- file ---\" markers; else \"--- page N ---\" + images\u002F\n$S find    \u003Ctopic words> [--sources youtube,arxiv,papers,hn,github] [--limit 8]   # candidate URLs, one block per platform\n$S pull    [--sources inbox,browser,ytwatchlater,ytliked] [--limit 10]   # ingest what the user saved or kept open; one line per new item\n```\n\nOutput goes to `~\u002F.cache\u002Fingest-url\u002F\u003Chash of url>\u002F` and a URL already ingested returns instantly\nas `cached`. Pass `--out \u003Cdir>` only when the user wants the files somewhere specific. The summary\nline names the directory and, for video, lists the chapters. `text.md` starts with front-matter\n(title, source, author, published) that Obsidian and graph tools read as is.\n\n## Finding sources, then ingesting them\n\n`find` returns one list per platform; views, citations, and points are not comparable, so you\nrank. Keyless: yt-dlp search, arXiv, Semantic Scholar with OpenAlex fallback, Hacker News, GitHub.\nX and Instagram have no free search; use the user's logged-in browser and say so. For \"what do talks on X say\" or\n\"what are people saying about X\": run `find`, pick the two or three hits that match best, ingest\nthose (`video` for talks, `article` for Hacker News threads, which come back as the full comment\ntree), and answer from them. Widen only if the user asks or the first pass disagrees\nwith itself. Do not fan out subagents over every hit: measured, one pass over three sources\ncosts about $1, a fan-out $7 and it hit the turn cap. General web questions belong to the\nharness's own web search.\n\n## Filing what you pulled\n\n`pull` ingests what the user saved on their phone or in YouTube and prints one line per new item\nwith its cache directory. The user's projects each have a status file at `~\u002Fnotes\u002F\u003Crepo>\u002FSTATUS.md`.\nFor every new item: read all status files, read the head of the item's `text.md`, and append one\nline to the one project it serves:\n\n    - 2026-09-10 | \u003Ctitle> | \u003Curl> | must-read | \u003Cone line: which open loop or next step it informs>\n\nThe grade is must-read, skim, or skip: must-read changes a decision in that project, skim adds\nbackground, skip is filed only so the pull is not repeated.\n\nin `~\u002Fnotes\u002F\u003Crepo>\u002FSOURCES.md`. If it serves none, append the same line to `~\u002Fnotes\u002Finbox\u002FUNSORTED.md`\nwith the reason it fits nothing. Never file without a stated reason, never file into two projects,\nand never invent a project. Unreachable or empty items are already reported by `pull`; leave them.\n\n## Reading what you ingested\n\n1. Map first. Video: the chapter list is already in the summary line. Paper: `grep -n '^--- \\|^#\\|\\\\\\\\section' text.md`.\n2. Targeted question: `grep -n -i \u003Cterm> text.md`, then read a window around each hit (two minutes\n   of `[mm:ss]` lines, or one section). Answer with quoted lines and their timestamps or pages.\n   Do not read the whole file for a fact; focused context is more reliable than full context.\n3. Whole-source summary: read `text.md` end to end, quote the lines you will rely on first, then\n   synthesize. For \"every number\" or \"all recommendations\", summarize per chapter or section and merge.\n4. Pull a frame only when the transcript points at something visual (\"as you can see\", \"this\n   chart\", a figure that is not spoken). One `ffmpeg -ss` frame beats a sheet survey.\n5. Every number you cite must be grep-able in `text.md` or visible in a named frame. If it is not\n   there, say so; do not fill the gap from memory.\n6. As a subagent, return quotes with timestamps plus the directory path, not the transcript.\n\n## Pick the cheapest path that answers the question\n\n- Title, chapters, duration only: `video` (default) already prints them in its summary line and is\n  cached, so it costs nothing on repeat. Do not shell out to yt-dlp or ffmpeg yourself.\n- What was said: `video` (default). Read `text.md`: `CHAPTER mm:ss title` lines, then one\n  `[mm:ss]` line per minute. Cite these timestamps.\n- One slide or moment at a known time: run `video --frames` once, then open the one\n  `frames\u002FSSSSS.jpg` nearest that second.\n- Survey every slide or demo in a talk: `--frames`. Read `sheets\u002FNN.jpg` first (48 frames\n  per sheet, each labelled mm:ss, about 10 sheets per hour), then open only the\n  `frames\u002FSSSSS.jpg` you need. Whole-video frames are the expensive path; take it last.\n\n## When the script says no\n\n- `no text extracted` or `fetch failed` on an article: paywall or client-rendered page. Open it\n  with a browser tool that carries the user's login. Same for X posts and Instagram captions.\n- yt-dlp 403\u002F429: the script retries once with browser cookies and Chrome impersonation, and says\n  so. If it still fails, the site changed: run `uv tool upgrade yt-dlp`, retry once, then report.\n- `speech-to-text returned no transcript`: the first no-caption video installs `scripts\u002Fstt.py`'s\n  environment (about 2 GB); if that is unwanted, say so instead of retrying.\n- The first call on a machine installs the environment (about 10 s); do not treat that wait as a failure.\n- Do not \"succeed\" by summarizing from the title, description, or memory. Report the failure.\n\nDesign choices and measurements: `EVAL.md`. Self-check: `python3 scripts\u002Ftest_ingest.py`.\n","# Ingest URL\n\nOne command per source type. The script writes a directory, prints a one-line summary ending in\nwhat to read next, and on any failure exits nonzero with the reason. It never writes empty text.\nPaths below are relative to this skill's base directory.\n\n```bash\nS=\u003Cbase directory>\u002Fscripts\u002Fingest.py\n$S video   \u003Curl> [\u003Curl> ...]          # text.md + manifest.json per URL, ~15s each. Captions, else speech-to-text\n$S video   \u003Curl> --frames             # also source.mp4, frames\u002FSSSSS.jpg (name = seconds), sheets\u002FNN.jpg\n$S article \u003Curl> [\u003Curl> ...]          # text.md with front-matter; HN item URLs become the comment tree\n$S pdf     \u003Curl-or-path> [\u003Curl> ...]  # arXiv: LaTeX source with \"--- file ---\" markers; else \"--- page N ---\" + images\u002F\n$S find    \u003Ctopic words> [--sources youtube,arxiv,papers,hn,github] [--limit 8]   # candidate URLs, one block per platform\n$S pull    [--sources inbox,browser,ytwatchlater,ytliked] [--limit 10]   # ingest what the user saved or kept open; one line per new item\n```\n\nOutput goes to `~\u002F.cache\u002Fingest-url\u002F\u003Chash of url>\u002F` and a URL already ingested returns instantly\nas `cached`. Pass `--out \u003Cdir>` only when the user wants the files somewhere specific. The summary\nline names the directory and, for video, lists the chapters. `text.md` starts with front-matter\n(title, source, author, published) that Obsidian and graph tools read as is.\n\n## Finding sources, then ingesting them\n\n`find` returns one list per platform; views, citations, and points are not comparable, so you\nrank. Keyless: yt-dlp search, arXiv, Semantic Scholar with OpenAlex fallback, Hacker News, GitHub.\nX and Instagram have no free search; use the user's logged-in browser and say so. For \"what do talks on X say\" or\n\"what are people saying about X\": run `find`, pick the two or three hits that match best, ingest\nthose (`video` for talks, `article` for Hacker News threads, which come back as the full comment\ntree), and answer from them. Widen only if the user asks or the first pass disagrees\nwith itself. Do not fan out subagents over every hit: measured, one pass over three sources\ncosts about $1, a fan-out $7 and it hit the turn cap. General web questions belong to the\nharness's own web search.\n\n## Filing what you pulled\n\n`pull` ingests what the user saved on their phone or in YouTube and prints one line per new item\nwith its cache directory. The user's projects each have a status file at `~\u002Fnotes\u002F\u003Crepo>\u002FSTATUS.md`.\nFor every new item: read all status files, read the head of the item's `text.md`, and append one\nline to the one project it serves:\n\n    - 2026-09-10 | \u003Ctitle> | \u003Curl> | must-read | \u003Cone line: which open loop or next step it informs>\n\nThe grade is must-read, skim, or skip: must-read changes a decision in that project, skim adds\nbackground, skip is filed only so the pull is not repeated.\n\nin `~\u002Fnotes\u002F\u003Crepo>\u002FSOURCES.md`. If it serves none, append the same line to `~\u002Fnotes\u002Finbox\u002FUNSORTED.md`\nwith the reason it fits nothing. Never file without a stated reason, never file into two projects,\nand never invent a project. Unreachable or empty items are already reported by `pull`; leave them.\n\n## Reading what you ingested\n\n1. Map first. Video: the chapter list is already in the summary line. Paper: `grep -n '^--- \\|^#\\|\\\\\\\\section' text.md`.\n2. Targeted question: `grep -n -i \u003Cterm> text.md`, then read a window around each hit (two minutes\n   of `[mm:ss]` lines, or one section). Answer with quoted lines and their timestamps or pages.\n   Do not read the whole file for a fact; focused context is more reliable than full context.\n3. Whole-source summary: read `text.md` end to end, quote the lines you will rely on first, then\n   synthesize. For \"every number\" or \"all recommendations\", summarize per chapter or section and merge.\n4. Pull a frame only when the transcript points at something visual (\"as you can see\", \"this\n   chart\", a figure that is not spoken). One `ffmpeg -ss` frame beats a sheet survey.\n5. Every number you cite must be grep-able in `text.md` or visible in a named frame. If it is not\n   there, say so; do not fill the gap from memory.\n6. As a subagent, return quotes with timestamps plus the directory path, not the transcript.\n\n## Pick the cheapest path that answers the question\n\n- Title, chapters, duration only: `video` (default) already prints them in its summary line and is\n  cached, so it costs nothing on repeat. Do not shell out to yt-dlp or ffmpeg yourself.\n- What was said: `video` (default). Read `text.md`: `CHAPTER mm:ss title` lines, then one\n  `[mm:ss]` line per minute. Cite these timestamps.\n- One slide or moment at a known time: run `video --frames` once, then open the one\n  `frames\u002FSSSSS.jpg` nearest that second.\n- Survey every slide or demo in a talk: `--frames`. Read `sheets\u002FNN.jpg` first (48 frames\n  per sheet, each labelled mm:ss, about 10 sheets per hour), then open only the\n  `frames\u002FSSSSS.jpg` you need. Whole-video frames are the expensive path; take it last.\n\n## When the script says no\n\n- `no text extracted` or `fetch failed` on an article: paywall or client-rendered page. Open it\n  with a browser tool that carries the user's login. Same for X posts and Instagram captions.\n- yt-dlp 403\u002F429: the script retries once with browser cookies and Chrome impersonation, and says\n  so. If it still fails, the site changed: run `uv tool upgrade yt-dlp`, retry once, then report.\n- `speech-to-text returned no transcript`: the first no-caption video installs `scripts\u002Fstt.py`'s\n  environment (about 2 GB); if that is unwanted, say so instead of retrying.\n- The first call on a machine installs the environment (about 10 s); do not treat that wait as a failure.\n- Do not \"succeed\" by summarizing from the title, description, or memory. Report the failure.\n\nDesign choices and measurements: `EVAL.md`. Self-check: `python3 scripts\u002Ftest_ingest.py`.\n",{"name":5,"license":15,"metadata":24,"description":7,"compatibility":28},{"author":25,"version":26,"verified":27},"Satyajeet Das","1.8.0","2026-09-08","macOS or Linux with uv and ffmpeg. Browser cookies (Chrome by default) are read only when a site blocks the plain request. Speech-to-text runs on-device in a separate environment installed on first need. Network required.","main","7c958155761b7fd0b98456012097c0974c010e3b","skills\u002Fingest-url",0,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url","https:\u002F\u002Fgithub.com\u002F123Satyajeet123","Hi👋👋\r\nThis is Satyajeet a student at IIT Bombay.\r\nInterested in web development AI, and entrepreneurship....","2026-09-15 23:14:34",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add 123Satyajeet123\u002Fingest-url --skill ingest-url",{"tool":43,"command":44},"bunx","bunx skills add 123Satyajeet123\u002Fingest-url --skill ingest-url",{"tool":46,"command":47},"pnpm","pnpm dlx skills add 123Satyajeet123\u002Fingest-url --skill ingest-url","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002FSKILL.md",true,[51,57,63,67,71,75],{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"EVAL.md","REFERENCE","参考文档",7260,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002FEVAL.md",{"path":58,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":61,"sourceHtmlUrl":62},"scripts\u002Fingest.py","SCRIPT","脚本",31617,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002Fscripts\u002Fingest.py",{"path":64,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":65,"sourceHtmlUrl":66},"scripts\u002Fpdf.py",876,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002Fscripts\u002Fpdf.py",{"path":68,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":69,"sourceHtmlUrl":70},"scripts\u002Fsetup",1023,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002Fscripts\u002Fsetup",{"path":72,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":73,"sourceHtmlUrl":74},"scripts\u002Fstt.py",1726,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002Fscripts\u002Fstt.py",{"path":76,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":77,"sourceHtmlUrl":78},"scripts\u002Ftest_ingest.py",17253,"https:\u002F\u002Fgithub.com\u002F123Satyajeet123\u002Fingest-url\u002Fblob\u002Fmain\u002Fskills\u002Fingest-url\u002Fscripts\u002Ftest_ingest.py",[]]