read-later-skill

name	read-later-skill
description	Use when the user sends article links in a chat or agent environment for save-for-later. Summarize and classify the article, store it in MongoDB via MongoDB MCP, and answer later questions with evidence and the source link.

Read Later Skill

When to use

Use this skill when the user wants to:

• save links for later reading,
• get concise summaries and tags,
• ask follow-up questions like "what did that article say?",
• retrieve original links quickly.

Preconditions

• MongoDB MCP is connected and write-enabled.
• Target database and collection are fixed:
• Database: readlater
• Collection: articles
• The collection follows references/mongodb-collection-schema.md.

Core workflow

1. Receive link + optional user note from the current chat or agent environment.
2. Normalize URL (remove obvious tracking params).
3. Build url_hash from normalized URL.
4. Deduplicate:

• First by canonical_url.
• Fallback by url_hash.

5. Build structured document:

• title,
• source site,
• user note,
• short summary,
• key points,
• tags,
• topic,
• language,
• saved time.

6. Write to MongoDB:

• Insert if not found.
• Else update only whitelisted fields.

7. Return acknowledgment:

• one-sentence summary,
• tags,
• record id,
• original URL.

Retrieval workflow

When user asks questions (for example "last week the xxx article said what?"):

1. Parse constraints:

• time range (for "last week", use the previous Monday-Sunday in user locale),
• topic keywords,
• optional source/site.

2. Query MongoDB documents with filters.
3. Rank candidates by topic/tag match and recency.
4. Answer with evidence-first format:

• short answer,
• 2-4 bullet key points,
• original article URL,
• record id.

5. If confidence is low, ask one clarification question before final answer.

Response contract

Always include:

• Answer: direct response in plain language.
• Why: evidence from stored summary/key points.
• Links: original URL first, record id second.

Never invent article details. If record data is insufficient, say so explicitly.

Write policy

• Keep summary to 120-220 Chinese characters by default.
• Keep key points to 3-5 bullets.
• tags: generate 3-6 specific tags from the article content. Tags are free-form and should not be limited to the topic taxonomy.
• topic: choose one primary category from this stable taxonomy:
• AI
• Engineering
• Product
• Business
• Design
• Data
• Science
• Media
• Society
• Health
• Lifestyle

Read policy

• For "what did it say", prefer stored Key Points and Evidence Snippets.
• For "exact wording", return best available snippet and source link.
• For "which one was that article", return top 3 candidates with links.

Safety rules

• Never write outside readlater.articles.
• Never upsert without dedupe check.
• Only update these fields:
• title
• summary
• key_points
• tags
• topic
• user_note
• last_updated_at
• If MongoDB MCP is unavailable, return an explicit error and ask for MCP fix. Do not fallback to local memory files.

References

• Collection schema: references/mongodb-collection-schema.md
• Answer templates: references/reply-templates.md
• Smoke test: references/smoke-test.md