Session Debrief

Session Debrief - Workshop Granola -> Transcript Summary + Q&A Update

You convert a completed NextLeap Applied Generative AI Bootcamp session into two artifacts:

1. A session summary at .local/session-transcripts/may-2026/session-<N>-<topic>-summary.md (folder month may differ - match the host's current cohort window).
2. Appended Q&A entries in 02-presentation/q&a.md, continuing the existing question numbering.

When this skill fires

The user gives you a Granola link (notes.granola.ai/t/...) with a phrase like "session completed", "process this session", "update transcripts and q&a", or just the link with implicit context that they just finished teaching.

Required inputs

• Granola UUID - extracted from the URL (the part after /t/ and before any suffix). Example: https://notes.granola.ai/t/9fb2ae8d-e2aa-415a-9518-9b2bd9dbeca3-008umkv4 -> UUID is 9fb2ae8d-e2aa-415a-9518-9b2bd9dbeca3.
• Today's date - use the workshop's fictional "now" if the project is anchored to one (per CLAUDE.md -> ## Time Convention). Otherwise use currentDate from system context. Always write dates as DD-MM-YYYY.

Steps

1. Fetch the transcript

Use the Granola MCP tool to load the transcript:

mcp__claude_ai_Granola__get_meeting_transcript(meeting_id=<UUID>)

The result is usually 100k+ chars - it will be saved to a tool-results file. Read it in chunks with Python (open(path).read()[start:end]) rather than the Read tool, since lines are very long.

2. Locate the existing session structure

Check both files to learn the format you must match:

• .local/session-transcripts/<month>-2026/ - look at the most recent session-N-<topic>-summary.md to mirror structure exactly. As of the June 2026 cohort, the established section order is: title, metadata (date, cohort, Granola link), Topics Covered, Audience Profile, Questions Asked (Key Themes), What Went Well, What Didn't Work, Key Quotes, Action Items Generated, an optional session-specific section if the call warrants one (e.g. a coverage/status recap), then Session Rating: X / 10 with a Why this rating - logical reasons breakdown. Always check the single most recent file in the most recent month folder - older files (pre-June 2026) used a leaner Demos Run Live / Open Items / Followups pattern that has since been superseded; don't mirror that older pattern.
• 02-presentation/q&a.md - read the date-context header block at the top to learn the current "Last full refresh" / "Additions" log, and skim the last "Session N Additions" block to learn:
  • The last Q number used (so the new questions continue numbering).
  • The exact heading format (e.g. # May 2026 - Session N Additions).
  • The per-question format - short answer, body, Sources: block with verified URLs.

3. Decide session number and topic

• Session number = last session number + 1.
• Topic = the main subject of the call (e.g. "multi-agents", "prototypes", "skills"). Derive from the first 5-10k chars of the transcript.

4. Write the session summary

File path: .local/session-transcripts/<current-month>-2026/session-<N>-<topic>-summary.md

Mirror the previous session's structure. Each section's quality bar:

• Topics Covered - 15-25 numbered items, each one a specific moment from the call, not a generic bullet. Use the host's actual phrasing where memorable. Fold live demos into the relevant numbered item (bold the demo name) rather than a separate "Demos Run Live" section - that matches current practice.
• Audience Profile - participant count, names if mentioned, behavioral observations (engagement, languages, who asked what kind of question), recurring themes.
• Questions Asked (Key Themes) - cluster by theme, quote questions verbatim where possible. This is the source of truth for the Q&A update below.
• What Went Well - 4-6 bullets on demos/moments that landed, concrete and specific (name the artifact or exchange, not just "good engagement").
• What Didn't Work - 3-6 bullets on friction, unresolved bugs, unclear outcomes, or content that didn't get covered. Bold the headline of each bullet.
• Key Quotes - split into Aman: and Participants: subsections, verbatim quotes pulled from the transcript (attribute participants by name where known, else "a participant").
• Action Items Generated - bullet list, prefixed by owner where known (e.g. "Kapil: ...", "All participants: ...", "Aman: ...").
• Session Rating: X / 10 - one paragraph justifying the score, then a ### Why this rating - logical reasons subsection with three bolded groups: What pulled the score up (+), What kept it from a 9 or 10 (-), and What would push this to a 9+ next time, each 3-5 bullets.
• Add a session-specific section between Action Items and Session Rating only if the call warrants it (e.g. a "what got built vs. still pending" recap for a multi-part build session) - skip it otherwise.

Length target: 110-190 lines (matches the last 10+ sessions in practice).

5. Append to q&a.md

Update the date-context header at the top:

- **May 2026 Session N additions:** DD-MM-YYYY (Q<start>-Q<end> - see "May 2026 - Session N Additions" section)

Then append a new section at the bottom of the file:

---

# May 2026 - Session N Additions

New questions raised by the NextLeap Applied Generative AI Bootcamp cohort on DD-MM-YYYY during Session N (<topic>). All URLs verified DD-MM-YYYY.

---

## Q<n>: <question verbatim or tightly paraphrased>

**Short answer:** <one-or-two-sentence direct answer>

<body - tables, lists, code blocks as needed. Cite specific docs, commands, and limits>

**Sources:**
- <url> (verified DD-MM-YYYY)
APPEND_GUARD

Convert every "Question Asked" cluster from the session summary into a numbered Q. Aim for 10-15 Qs per session (consolidate near-duplicates).

6. Quality checks before returning

• Session number continues from prior session (no gap, no overlap)
• Q numbering continues from the last Q in q&a.md (no gap)
• Every URL is validated via curl -s -o /dev/null -w "%{http_code}" -L --max-time 10 -A "Mozilla/5.0" <url> (per global CLAUDE.md rule). Replace any 404s before saving
• All dates are DD-MM-YYYY. No ISO, no US format
• No em dashes (per global writing rules)
• No emojis unless the user asked
• Date in q&a.md header block updated with the new session row
• File written under .local/session-transcripts/<month>-2026/, not outputs/ or elsewhere
• Granola link copied into the summary's metadata block
• Sources are from code.claude.com, docs.anthropic.com, or anthropic.com/engineering only - no third-party blogs unless the user explicitly asked

7. URL hygiene (non-negotiable)

This skill writes a lot of citation URLs. Per global rules:

• HEAD/GET-check every URL before writing it. Accept 2xx; LinkedIn/anti-bot 999 is OK.
• Beware soft-404s on docs.anthropic.com and code.claude.com - fetch body, grep for "404" / "Page not found".
• Use official sources only. If no first-party page exists for a topic, omit the citation rather than substitute a blog.

Output

End with a one-line summary like:

Wrote session-<N>-<topic>-summary.md and appended Q-Q to q&a.md. URLs verified.

Nothing else. The user is busy and just wants confirmation the artifacts are in place.

Common pitfalls

• Reading the Granola transcript with the Read tool. The lines are too long. Use Python char-range slicing instead.
• Inventing questions that weren't asked. Stay grounded in the transcript. If a Q is implied but not asked verbatim, mark it as a paraphrase in the heading.
• Drifting from the established structure. The host's q&a.md has a strict format developed over 5 sessions. Mirror it exactly - don't introduce new section names or reorder.
• Mirroring a stale session-summary template. The session-summary section list evolved (see step 2/4) - always check the single most recent file in the most recent month folder, not an older one, and don't skip What Went Well / What Didn't Work / Key Quotes / Action Items Generated / Session Rating just because an older file in a different month omitted them.
• Wrong month folder. Check the existing .local/session-transcripts/ folder names; don't create a new month folder unless the previous session's month no longer applies.
• Citing third-party content. Stick to official Anthropic / Claude docs unless the user explicitly asked otherwise.