glebis/claude-skills/coaching-session-summarizer

Coaching Session Summarizer

Overview

Analyzes a coaching/therapy session transcript and appends a structured summary (key insights, decisions, action items, deep analysis, connected trails) to the note.

The agent (Claude Code) performs the analysis directly — reading the transcript and writing the summary in this session. There is no Anthropic API call and no billing; it runs entirely on the active subscription. A legacy API-based script is kept only as a headless fallback (see bottom).

When to Use This Skill

• A new Fathom/Granola transcript was synced to the vault (coaching or therapy)
• User asks to summarize/analyze a session (/summarize-session [file] or similar)
• After calendar-sync or a Granola export, when a new *-coaching.md,

*-therapy.md, or *-session.md file appears — offer to summarize it

Workflow (agent-driven — default)

Do this in-session with native tools. No API key required.

Step 1 — Gather context

Run the deterministic helper to get the transcript text, previous sessions, and the trail list in one shot:

python3 ~/.claude/skills/coaching-session-summarizer/scripts/gather_context.py \
  <transcript-file> --vault ~/Brains/brain

It prints:

• Previous sessions with the same participant (paths) — Read these only in

deep mode, for cross-session pattern detection

• Available trails — pick 2–4 most relevant to link
• Session content — the summary + transcript to analyze (any prior

AI-Generated Summary is stripped so re-runs stay clean)

Pass --participant <name-slug> if the filename doesn't encode the person (e.g. Granola exports titled by topic): --participant gleb-kalinin.

Step 2 — Analyze

Read the session content and extract, in the analytical voice of a session analyst (objective, using the speaker's authentic language where it matters):

• Key Insights — 3–5 main realizations / breakthroughs / observations
• Decisions Made — concrete choices or commitments
• Action Items — specific next steps; prefix time-sensitive ones with

[URGENT] and scheduling items with [SCHEDULING]

• Session Themes — 2–3 recurring topics or patterns

Deep mode (default for therapy and milestone sessions) — also Read the previous sessions and add:

• Pattern Detection — themes recurring across sessions
• Progress Assessment — movement on earlier commitments
• Energy/Motivation Markers — shifts in energy, resistance, affect
• Potential Obstacles — what might block progress

Step 3 — Append with Edit

Append the summary to the end of the transcript file using Edit (never overwrite existing content). Match this exact structure:

## AI-Generated Summary
 
*Generated: YYYY-MM-DD*
 
### Key Insights
-...
 
### Decisions Made
-...
 
### Action Items
-[URGENT] ...
-...
 
### Session Themes
-...
 
## Deep Analysis
 
-**Pattern Detection**: ...
-**Progress Assessment**: ...
-**Energy/Motivation Markers**: ...
-**Potential Obstacles**: ...
 
## Connected Trails
 
-[[Trails/Trail - <Name>|<Name>]]
-[[Trails/Trail - <Name>|<Name>]]

Use the current date (date +%Y-%m-%d) in the Generated line. Omit the Deep Analysis section in quick mode. Verify trail link names against the printed trail list — case and exact wording matter for Obsidian links.

Modes

• quick — Key Insights, Decisions, Action Items, Themes. Skip Deep Analysis

and previous-session reads.

• deep (recommended for therapy / milestones) — everything, including

reading previous sessions for pattern detection.

Integration with Sync

After calendar-sync or a Granola/Fathom export, check for new session files (*-coaching.md, *-therapy.md, *-session.md). If one appears, offer: "New session detected — summarize now?" Default to deep mode for therapy.

Notes

• Preserves the original transcript intact; the summary is always appended.
• Trail linking requires the Trails/ directory in the vault root.
• Cross-session comparison works best with consistent naming:

YYYYMMDD-name-coaching.md / YYYYMMDD-name-therapy.md.

• Re-running is safe: gather_context.py strips any prior AI-Generated Summary

before printing, so the agent analyzes only the raw session. (Delete the old ## AI-Generated Summary block from the file before re-appending if you want to replace rather than stack summaries.)

Resources

scripts/

• gather_context.py — (default path) deterministic context gatherer, no

API. Prints transcript text + previous sessions + trail list for the agent to analyze in-session.

• summarize_session.py — legacy / headless fallback. Calls the Anthropic

API directly (model via SUMMARIZER_MODEL, default claude-sonnet-4-6) and bills a funded ANTHROPIC_API_KEY. Use only when no interactive agent is available (e.g. cron). Exits with a clear message if the key has no credit.