fix: improve What field quality and prevent duplicate entries

[StartupBros]

Problem

The **What:** field in bookmark entries frequently degrades to low-quality output across three failure modes:

1. Category labels instead of descriptions — "Tool or resource share", "Commentary/perspective", "Claude Code insights/comparison"
2. Echo of tweet text — The model parrots the tweet verbatim instead of synthesizing ("karpathy really is the fucking goat", "Are you guys starting to catch on?")
3. Empty or placeholder content — "Video content post", "Social media image bookmark", raw t.co URLs, or completely empty fields

This happens because the prompt gives almost no guidance on What field quality — just {1-2 sentence description of what this actually is}.

Solution

Three targeted improvements to process-bookmarks.md:

1. What Field Rules section

• Explicit quality minimum (80+ characters)
• NEVER-write list for common failure patterns
• THIN: prefix convention for image/video-only tweets where context is genuinely missing
• LINK_FAILED: prefix for unresolvable t.co links
• Special guidance for quote tweets (synthesize BOTH the reaction AND quoted content)

2. Subagent template quality rules

Subagents don't see the full prompt, so the What field rules are inlined in the subagent prompt template.

3. Deduplication check

Prevents duplicate entries when bookmarks are re-processed (e.g., after a --force fetch).

4. Better title guidance

• Quote tweets: use the quoted content's substance, not the reaction text
• Media-only posts: [Media] prefix instead of "Video post" or raw URLs

Testing

Tested on 419 bookmarks processed with both Haiku and Sonnet:

Metric	Before	After
Empty What fields	9 (2.1%)	0
Very Short (<50 chars)	37 (8.8%)	0
LINK_FAILED (properly tagged)	0	7
THIN (properly tagged)	0	4

Zero garbage descriptions after the fix.

🤖 Generated with Claude Code