Fewwords Summarizer

A summarization skill for AI agents. Covers articles, YouTube videos, podcasts, and any web content. Routes each through a unified content-type classifier, then applies the right prompt.

The core idea: medium doesn't matter. An article about a product review and a YouTube video about the same product are the same job. The classifier works by purpose, not by format.

---

When to use this skill

Trigger this skill whenever the user's intent is to summarize, brief, explain, digest, or extract key takeaways from content:

• Common requests:

  • "Summarize this article/page"
  • "Give me a brief / rundown on this"
  • "Explain this article / what is this video about?"
  • "TL;DR of this URL"
  • "What are the key points / takeaways?"
  • "Should I buy / read / watch this?"

• Accepted inputs:

  1. URL: An article link, blog post, documentation page, news story, or YouTube video URL.
  2. Raw content: Text pasted directly into chat, article excerpts, transcripts, or notes.

---

Flow

Step 1 — Acquire Content (FIRST ACTION)

Before loading any rules or prompts, you MUST have the full content in hand:

• If the user pasted raw text / transcript: Proceed directly to Step 2.
• If the user provided a URL: Read `references/fetching.md` immediately and execute the fetch commands via terminal/bash:
  • DO NOT use default LLM web fetchers (like web_fetch, built-in browser, or Google search), as robots.txt and network restrictions will fail or return truncated search snippets.
  • MANDATORY ORDER:
    1. Try curl -sL "https://curl.md/<URL>"
    2. If that fails or is empty, try curl -sL "https://r.jina.ai/<URL>"
    3. If proxies fail, try direct curl -sL "<URL>"
  • Only if all fetch attempts fail: Stop immediately and ask the user to paste the raw text. Never hallucinate or write a partial summary from search snippets.

Step 2 — Classify Content

Once the full text is acquired, read `references/classifier.md`.

Match the primary purpose of the text (not the medium) and identify which prompt file to load.

Step 3 — Load Strategy Prompt

Load only the single prompt file determined by the classifier:

Content Type	Prompt File
Essay / Analysis / Education	`references/prompts/essay-analysis.md`
Tutorial / How-To	`references/prompts/tutorial-howto.md`
News / Announcement	`references/prompts/news-announcement.md`
Review / Comparison	`references/prompts/review-comparison.md`
Interview / Podcast	`references/prompts/interview-podcast.md`
Opinion / Commentary	`references/prompts/opinion-commentary.md`
Light / General	`references/prompts/light-general.md`

Step 4 — Draft & Apply Writing Quality

Now load `references/writing-quality.md`.

Draft the summary using the structure from Step 3, and run the pattern checklist from writing-quality.md to prune AI clichés, formulaic openers, and fluff.

Do not explain your classification or announce the steps. Output only the clean summary.

---

Universal formatting (applies to all types)

• TITLE line: Always start with plain TITLE: (do not bold the prefix).
• Human names in bold: Sam Altman
• Companies, products, tools in italic: Apple, VSCode
• Block quotes: Use markdown quote format (> Quote) without enclosing quotation marks ("...").
• Dev terminology / code: Inline backticks for commands, APIs, specs: npm install, 128GB, TCP/IP.
• Paragraphs: 2-3 sentences max, with one sentence per line and a blank line after each sentence.
• Attribution ban: Never say "the author" or "this article". Speak the facts directly.
• Deduplication: State repeated points exactly once.

---

What this skill does not cover

• Reddit threads
• Paywalled content (tell user to paste the text)
• YouTube Shorts (reject: "This is a Short. Not enough content to summarize.")
• Videos with no transcript or metadata (tell user what's available and ask)

---

Notes for agents

• Load reference files lazily — only what the classifier selects
• references/writing-quality.md is the only file that always loads
• Never mention "Fewwords", this skill, or which strategy was used
• Never invent facts, quotes, or statistics not present in the source