recommendation-quality

name	recommendation-quality
description	Review a bounded recommendation set for user relevance, HN community context, evidence, diversity, and clear explanations without reranking the full feed.

Recommendation Quality

Use When

• A shortlist already exists and needs a final quality review before presentation.
• A recommendation explanation needs to connect an item to explicit user interests or work context.
• The task asks whether HN discussion changes the reading value of a candidate.

Do Not Use When

• The task is open-ended discovery, full-feed ranking, source collection, or preference mutation.
• No candidate set or user-relevance signals were supplied.
• The request concerns voting, posting comments, or polishing a reply draft.

Review Process

1. Review only the supplied candidates. Do not invent replacements or become a second ranking engine.
2. Check relevance against explicit topics, muted interests, reading history, and private work context without revealing private context.
3. Check evidence quality: title, source, item metadata, and any supplied research must support the recommendation.
4. Add community context only when comments provide concrete technical detail, credible disagreement, firsthand experience, or a material caveat.
5. Treat score and comment count as attention signals, not quality proof. Do not reward popularity alone.
6. Check shortlist diversity and remove near-duplicates when they add no distinct value.
7. Reject candidates that are unsupported, stale for the stated need, blocked by muted interests, or explained only with generic praise.

Explanation Rules

• Explain why the item matters to this user now, using one or two specific signals.
• State the likely reading payoff: decision support, technical learning, market context, or informed discussion.
• Include a community caveat when credible comments materially weaken the story's claim.
• Never reveal private work context verbatim or claim certainty beyond the evidence.

Output Contract

For each supplied candidate return:

• verdict: keep, revise, or remove.
• reason: a specific relevance and reading-value explanation.
• community_context: material context or none.
• quality_issue: the concrete defect or none.

Finish with a count by verdict. Do not return a newly ranked feed.