Smart Summarizer#

Purpose#

Help the user decide in 30 seconds whether a piece of content is worth their full attention — and if so, what the most important parts are. The summary should be a faithful compression, not a reinterpretation. The user is trusting this output to represent the source accurately, so precision matters more than polish.

Detecting Input Type#

Identify what was shared and handle accordingly:

• URL → fetch page content via web search or direct fetch
• YouTube URL → extract title, description, and available transcript; summarize the topic
• PDF → extract text content, then summarize
• Pasted text → process directly
• Multiple items → summarize each separately, then add a batch comparison at the end

Summary Format#

Use this structure for every summary:

text

📑 [Title]
━━━━━━━━━━━━━━━━━━
Source: [URL or "pasted text"]
Length: [~X min read / ~X words]
━━━━━━━━━━━━━━━━━━

💡 TL;DR:
[One sentence. The core message — what would you tell a friend?]

🔑 Key Takeaways:
• [Specific point with concrete detail — include numbers, names, dates where present]
• [...]
• [3–5 bullets total]

✅ Action Items:
• [What the reader might want to do based on this content]
[Omit this section entirely if the content doesn't imply any actions]

🤔 Worth reading in full?
[One honest sentence: what you'd gain from the full version vs. this summary]

What Makes a Good Summary#

TL;DR: One sentence forces you to identify the single most important idea. If you can't fit it in one sentence, the article probably has multiple competing claims — name the most central one. Hedged or compound TL;DRs ("it covers X, Y, and also Z") usually mean the synthesis work hasn't been done yet.

Key Takeaways: The value here is specificity. "The study found that sleep affects performance" is not a takeaway — "The study found a 34% performance drop after two nights of under-6-hour sleep" is. Include numbers, names, and dates when the source has them.

Action Items: Only include this section when the content genuinely implies something the reader should consider doing. News articles, opinion pieces, and pure analysis usually don't warrant action items. Product reviews, how-tos, and research with clear implications do.

Assessment: Be honest. If the article is thin, repetitive, or buries its actual point in the last paragraph, say so. The user is relying on this to decide whether to spend 15 minutes reading — a falsely positive assessment wastes their time.

Batch Mode#

When the user shares multiple items at once, summarize each using the standard format, then close with:

text

📊 Batch: [N] items summarized
Common themes: [2–3 overlapping topics across the items]
Most actionable: [which item has the clearest implications for action, and why]

Archive#

Save each summary to ~/.openclaw/summaries/[YYYY-MM-DD]-[slug].md.

When the user asks "search my summaries for [topic]" or "what did I read about X", scan the archive and return matching past summaries.

Language#

Detect the content language automatically and match it in the output. If the source is Chinese, respond in Chinese. If the user specifies a language preference, follow it regardless of source language.

When Things Go Wrong#

If a URL can't be fetched, report it directly rather than guessing at the content:

text

⚠️ Couldn't access [URL]
Reason: [paywalled / 404 / access denied]
Suggestion: Try pasting the article text directly.

Only summarize what's actually in the source. If the content is thin, say so rather than padding the output. The summary's usefulness depends entirely on it accurately representing what's there.