synthesize

name	synthesize
description	Use when processing pending session transcripts into daily summaries — first session of the day, on manual request, or on schedule
user-invokable	true

Synthesize Skill

Launch a subagent to process memory transcripts into daily summaries and selectively route key learnings to long-term memory.

Execution

IMPORTANT: This skill MUST be executed by launching a subagent. Do NOT run synthesis in the main context.

1. Get the synthesis prompt, model, and pre-extracted data:

   uv run --no-project $HOME/.claude/scripts/load_memory.py --synthesis-prompt
   

   • If output says "No pending transcripts", inform the user and stop.
   • First line: model=<model>. Second line: prompt_file=<path>.

2. Launch a synthesis subagent (foreground — manual /synthesize always blocks so user sees results). IMPORTANT: Do NOT read the prompt file. Pass the file path to the subagent and let it read the file itself. This avoids the main agent regenerating ~50K tokens of prompt content.

   Task(
     subagent_type: "general-purpose",
     model: <model from first line>,
     prompt: "Read <prompt_file path> and follow the instructions in it exactly. Use only Write and Bash tools."
   )
   

3. Report the subagent's summary to the user.

---

Reference: Tag Types

Actions: implement, improve, document, analyze Decisions: design, tradeoff, scope Learnings: gotcha, pitfall, pattern Lessons: insight, tip, workaround

Reference: Pinning Criteria

Items in long-term memory can be moved to ## Pinned section (protected from decay):

• Fundamental architecture patterns
• Safety-critical information
• Cross-project patterns that proved valuable over time