Feed bookmark / tag / note feedback into Skill synthesis as a human signal

[chigichan24]

Why

• The bookmarks, tags, and notes collected via Bookmark / tag / annotate valuable moments in Session Playback (human feedback channel) #23 are the strongest human signal we can gather about which moments are actually reusable.
• The current synthesis prompt in scripts/skill-synthesizer.ts already weaves in TF-IDF keywords and facets-derived summaries, but it never sees the moments humans explicitly flagged as valuable.

What

• Inject summaries of bookmarked turns into the Skill synthesis prompt (kept short — the goal is signal, not transcript replay).
• Use tag semantics:
  • reusable → boost the moment as positive evidence for the Skill candidate.
  • anti-pattern → include as a counter-example, with explicit guidance to avoid this pattern.
• Add a humanSignal term to the reusability score so bookmarked clusters rank higher.
• Use the evaluator from Add an automatic Skill evaluation pipeline (skill-creator-style quality checks) #20 to compare synthesized Skills with vs. without human-signal enrichment, so we can verify the loop actually improves quality.

How

• Wait for the persistence format from Bookmark / tag / annotate valuable moments in Session Playback (human feedback channel) #23 to land before extending buildSynthesisPrompt (or whatever it ends up being called).
• Keep the prompt extension behind a flag (--use-human-feedback) initially so we can A/B against the no-feedback baseline.

Depends on

• Bookmark / tag / annotate valuable moments in Session Playback (human feedback channel) #23 (bookmark / feedback capture)
• Add an automatic Skill evaluation pipeline (skill-creator-style quality checks) #20 (Skill evaluator — needed to measure the impact)