distill

name	distill
description	Distill insights from a experimentalist agent's session histories — zero API calls

Distill

Extract the essence from a experimentalist agent's work. Read its Claude Code session histories, cross-reference with runs.jsonl, and produce structured research insights.

Arguments

The user provides an experiment name or project ID (e.g. "tiny-matmul").

Steps

1. Resolve the project — call mcp__distillate__get_project_details to get the project path and run list.

2. Find session histories — Claude Code sessions are at ~/.claude/projects/ with a path-based key. Use Glob to find ~/.claude/projects/*<project-name>*/*.jsonl. Each .jsonl file is one experimentalist agent session.

3. Read sessions — for each session file (newest first, up to 10), read it and extract:

   • Agent reasoning — assistant messages with type: "text" blocks
   • Thinking blocks — type: "thinking" in assistant content
   • Tool calls — what the agent read, edited, ran
   • Run announcements — writes to runs.jsonl (status: "running", "keep", "discard")

4. Cross-reference with runs — match sessions to runs by timestamp overlap. For each run:

   • What hypothesis the agent was testing
   • What changes it made (from Edit/Write tool calls)
   • Why it kept or discarded the run
   • Key metrics achieved

5. Synthesize — across all sessions, identify:

   • Key breakthrough: the single most impactful discovery
   • Lessons learned: 3-5 actionable insights
   • Dead ends: approaches tried and abandoned
   • Trajectory: how the agent's strategy evolved

6. Save enrichment — call mcp__distillate__save_enrichment with the project name and structured insights. These appear in the desktop UI — write for scannability, not for a paper:

   • key_breakthrough: One sentence. State the metric improvement and what caused it. No Greek letters, no parenthetical asides, no compressed notation.
   • lessons_learned: 3-5 short sentences. Each starts with the finding, then one supporting number. No ALL CAPS. Write like you're explaining to a smart colleague.
   • dead_ends: One sentence each — name the approach and why it failed.
   • trajectory: 2-3 sentences — the story arc from baseline to current best.
   • run_insights: dict of per-run insights (keyed by run ID)

   This writes to .distillate/llm_enrichment.json and the insights immediately appear in the desktop Control Panel.

7. Report — summarize: sessions analyzed, runs enriched, key breakthrough.

Important

• Session files are JSONL — one JSON object per line. Use the Read tool.
• Focus on assistant messages — skip user messages (just tool results).
• Session dir path: /Users/foo/experiments/tiny-matmul → ~/.claude/projects/-Users-foo-experiments-tiny-matmul/
• Don't invent insights — only report what the agent wrote. Quote its words.
• This skill makes ZERO API calls. All data comes from local session files.