meeting-workshop

name	meeting-workshop
description	Passive live-meeting helper. Polls a Notion meeting-notes page (transcript) every minute via cron, surfaces only deltas, and proposes <USER_NAME>voice answers when a question or decision appears. Optionally watches a Google Doc's comment stream alongside. <USER_NAME>drives the meeting; this skill watches and feeds. Never auto-edits the doc. Use when <USER_NAME>says "/meeting-workshop", "monitor this meeting", "watch this transcript", "live meeting helper", or pastes a Notion meeting-notes URL during a call.
argument-hint	<notion-url> | <notion-url> <doc-id> | status | stop
category	execution
runtimes	["claude"]
pii_safe	true

Meeting-Workshop

Passive live-meeting helper. Watches a Notion meeting transcript and a Google Doc's comment stream concurrently, fires every minute, and surfaces only what's new. <USER_NAME>runs the meeting. The skill feeds him answers and decisions in real time so he never has to break flow to look things up.

Roles

• **<USER_NAME>is in the meeting (Discord/Zoom/in-person), driving the conversation and the doc edits.
• Claude (this skill) watches via mcporter on <INFERENCE_HOST>. No browser, no auto-edits. Surface deltas, propose <USER_NAME>voice answers, stop.

These roles are fixed.

When to use

• Live meeting with a customer / coworker over a Google Doc deliverable
• Notion is recording the transcript on a meeting-notes page
• <USER_NAME>wants real-time delta surfacing instead of post-hoc transcript review
• <USER_NAME>may want to land an edit during the meeting, but the skill never executes it without explicit approval

Do NOT use for:

• Async doc review (use pair-edit instead)
• Single-shot transcript summary (use /live <terms> instead)
• Anything requiring auto-execution of edits (this skill is read-only)

Inputs

/meeting-workshop <notion-url> — start monitoring on this Notion meeting page. Transcript-only mode. Use this as the default trigger when <USER_NAME>pastes a Notion meeting-notes URL during a live call.

/meeting-workshop <notion-url> <doc-id> — also monitor open comments on the named Google Doc.

/meeting-workshop status — show current cron job id, page, doc id, sweep count, last delta.

/meeting-workshop stop — cancel the cron and dump final state.

Recognizing the trigger: if <USER_NAME>pastes a notion.so/...Monthly-Sync... or any meeting-notes URL and says anything like "you are my live meeting helper", "check every minute", "watch this", or "answer questions as they come" — this is the skill. Start the cron immediately. Do NOT poll with a sleep/while loop, do NOT background bash with &. Use CronCreate. (<USER_NAME>got burned by backgrounded shells on 2026-05-14 — they fight the harness and the wake signal stops landing in chat.)

Setup (run once when the user invokes with a URL)

1. Reset state files

mkdir -p /tmp/meeting-workshop
> /tmp/meeting-workshop/last-transcript.txt
> /tmp/meeting-workshop/known-comments.txt
echo "<notion-url>" > /tmp/meeting-workshop/page.txt
echo "<doc-id>" > /tmp/meeting-workshop/doc.txt

2. Capture the baseline (so the first cron sweep doesn't dump the entire prior transcript as "new")

Run the same fetch the cron will run, save the result, and copy to last-transcript.txt / known-comments.txt. This step is the one that distinguishes a clean start from a noisy one.

3. Schedule the cron

Use CronCreate with cron: "* * * * *", recurring: true, and a prompt that calls back into this skill OR runs the sweep inline. The prompt body is the Sweep block below — embed the URL and doc id as literal strings inside the prompt so each fire has them.

Tell <USER_NAME>the cron job id and that recurring tasks auto-expire after 7 days.

Sweep block (this is what fires every minute)

NOTION_URL=$(cat /tmp/meeting-workshop/page.txt)
DOC_ID=$(cat /tmp/meeting-workshop/doc.txt)

ssh <INFERENCE_HOST> "mcporter call notion.notion-fetch \"id=$NOTION_URL\" \"include_transcript=true\"" \
  > /tmp/meeting-workshop/current-fetch.json 2>&1

ssh <INFERENCE_HOST> "mcporter call google-workspace.list_document_comments \"user_google_email=<EMAIL_ADDRESS>\" \"document_id=$DOC_ID\"" \
  > /tmp/meeting-workshop/current-comments.txt 2>&1

python3 << 'PYEOF'
import json, re, sys
with open('/tmp/meeting-workshop/current-fetch.json') as f:
    raw = re.sub(r'^Warning:[^\n]*\n', '', f.read())
try:
    text = json.loads(raw).get('text', '')
    m = re.search(r'<transcript>(.*?)</transcript>', text, re.DOTALL)
    transcript = m.group(1).strip() if m else ''
    with open('/tmp/meeting-workshop/current-transcript.txt', 'w') as f:
        f.write(transcript)
except Exception as e:
    print(f"FETCH-FAIL: {e}")
    sys.exit(1)
PYEOF

[ $? -ne 0 ] && { echo "skipping diff (Notion timeout — state preserved)"; exit 0; }

sed 's/^[\t ]*//' /tmp/meeting-workshop/current-transcript.txt > /tmp/meeting-workshop/current-norm.txt
sed 's/^[\t ]*//' /tmp/meeting-workshop/last-transcript.txt > /tmp/meeting-workshop/last-norm.txt

echo "---NEW TRANSCRIPT---"
comm -13 <(sort /tmp/meeting-workshop/last-norm.txt) <(sort /tmp/meeting-workshop/current-norm.txt)

echo "---NEW COMMENT IDS---"
grep -oE 'AAAB53Gd[A-Za-z0-9_-]+' /tmp/meeting-workshop/current-comments.txt | sort -u > /tmp/meeting-workshop/current-ids.txt
grep -oE 'AAAB53Gd[A-Za-z0-9_-]+' /tmp/meeting-workshop/known-comments.txt | sort -u > /tmp/meeting-workshop/prior-ids.txt
comm -23 /tmp/meeting-workshop/current-ids.txt /tmp/meeting-workshop/prior-ids.txt

cp /tmp/meeting-workshop/current-transcript.txt /tmp/meeting-workshop/last-transcript.txt
cp /tmp/meeting-workshop/current-comments.txt /tmp/meeting-workshop/known-comments.txt

After the sweep — Claude's response loop

When the cron fires and dumps a sweep result back to Claude, Claude inspects ---NEW TRANSCRIPT--- and ---NEW COMMENT IDS---:

If both empty

Output one line: no change. Stop. Do not summarize the standing transcript.

If new transcript content appears

Classify each chunk:

• Small talk / off-topic (family, pets, weather, politics, business gossip): one-line acknowledgment such as Still small talk (vet stories). No report content. Do not transcribe or summarize the small talk.
• Report-related question ("is the hash stored on the subscriber's machine or in Marathon"): quote the question verbatim, then propose a <USER_NAME>voice forensic answer using prior conversation context (<PROJECT_NAME> <PROJECT_NAME> report tabs, doc id from /tmp/meeting-workshop/doc.txt, feedback_lance_voice_formal.md).
• Decision ("let's say execution chain"): quote the decision verbatim, then state the action it implies (e.g., scrub doc for kill chain → execution chain). Do not auto-execute. Wait for <USER_NAME>to say "do it".
• Status update ("I condensed the executive summary to a one-pager"): one-line summary of what <TEAM_MEMBER> or the other party reported doing.

If new comment ids appear

Pull the comment content from /tmp/meeting-workshop/current-comments.txt for each new id. Quote the new content. If it's a <TEAM_MEMBER> comment that needs a <USER_NAME>voice reply, draft the reply but do not post it (replies are an explicit-approval action — see feedback_google_docs_suggesting_mode.md).

Cadence and tone

• One short paragraph or bullet list per sweep. Never long-form.
• Don't restate the standing context every sweep. <USER_NAME>has it loaded.
• If small talk runs for many sweeps, output no change or one-line "still small talk" acknowledgment. Don't pad.
• Notion timeouts are common (the MCP times out at 60 s sometimes). If FETCH-FAIL, the script preserves prior state. Output one line: Notion timeout. State preserved. Move on.

Voice register

Proposed answers for <USER_NAME>must follow feedback_lance_voice_formal.md:

• No em-dashes
• No analyst-tone leads ("Great question, …")
• No "you're right" / "good catch" / "good call" preamble
• Period-stop bullets
• Forensic register
• State the answer, not the meta

Hard rules

1. Never auto-edit the doc. This skill is read-only by contract. Edits are landed via pair-edit after the meeting, with explicit <USER_NAME>approval per feedback_google_docs_suggesting_mode.md.
2. Never resolve a comment automatically. Comments stay open per <USER_NAME>s standing rule unless he says to resolve.
3. Don't dump the whole transcript. Surface only deltas. The whole transcript is in Notion already.
4. Filter small talk as noise. One-line acknowledgment, then move on. Don't waste <USER_NAME>s chat scroll on dog stories.
5. Honor LIVE-REGISTER memory if no doc id is passed. Fall back gracefully if no memory exists; ask <USER_NAME>which doc to monitor.
6. Verify before claiming a delta. If the sweep returns no change, output no change and stop. Do not invent activity.
7. Notion is flaky. ~30% of fetches time out. The script must preserve prior state on timeout. Don't nuke last-transcript.txt on a failed fetch.
8. Stay one-shot per sweep. Do not chain multiple actions on a single fire. The cron will fire again next minute.

status subcommand

echo "page: $(cat /tmp/meeting-workshop/page.txt 2>/dev/null || echo 'NONE')"
echo "doc:  $(cat /tmp/meeting-workshop/doc.txt 2>/dev/null || echo 'NONE')"
echo "transcript: $(wc -c < /tmp/meeting-workshop/last-transcript.txt 2>/dev/null) chars"
echo "known comments: $(grep -oE 'AAAB53Gd[A-Za-z0-9_-]+' /tmp/meeting-workshop/known-comments.txt 2>/dev/null | sort -u | wc -l)"

Then call CronList to show the active cron job and its id.

stop subcommand

Call CronList. Find the cron job whose prompt mentions the meeting-workshop sweep block (or whose id matches the last /meeting-workshop run). Call CronDelete with that id. Output: Stopped meeting-workshop. Final transcript at /tmp/meeting-workshop/last-transcript.txt.

Anti-patterns

• "Let me transcribe the small talk so <USER_NAME>can scan it" — no, filter as noise.
• "Resolved the comment for <USER_NAME>no, comments stay open.
• "Already drafted the reply, posted it, here's the result" — no, draft only, never post.
• "Edited the doc to fix the kill-chain wording in real time" — no, editing is pair-edit's job and requires explicit approval.
• "Re-summarized the entire meeting context every sweep" — no, surface deltas, <USER_NAME>has the rest.
• "Sleep-polled instead of using the cron" — no, the cron is the wake signal.

Cross-reference

• pair-edit skill — the explicit-approval edit ritual to run AFTER the meeting on items the workshop surfaced
• feedback_lance_voice_formal.md — voice rules for proposed answers
• feedback_google_docs_suggesting_mode.md — why automated edits to deliverable docs need explicit approval
• /live skill — single-shot Notion fetch (use that for one-off queries; meeting-workshop is the recurring variant)
• Working examples:
  • 2026-05-07 14:22 EDT (<PROJECT_NAME> <PROJECT_NAME> report, Nyssa) and 17:10 EDT follow-up — ~70 sweeps, validated the deltas-only + <USER_NAME>voice-answer pattern.
  • 2026-05-14 12:22 EDT (<PROJECT_NAME> monthly sync, <EXTERNAL_PARTY> + Nyssa + <EXTERNAL_PARTY>) — ~57 sweeps, surfaced live: Havok+Wwise+Marathon prototypes in cheat module, lead-tester focus pivot, 1-3 person dev team confirmation, "send AOB framework to Chris" action item. Validated the trigger flow of <USER_NAME>pasting a notion.so meeting URL and saying "you are my live meeting helper."

Today's lessons (2026-05-14)

• The MCP tool returns {"text": "<meeting-notes>..."}. Extract via json.loads(raw).get('text',''), NOT .get('result',''). The result key is empty.
• include_transcript=true is mandatory. Without it, the response says "Transcript omitted. Use the view tool…" and the regex extraction returns empty.
• Notion timeouts at ~30% rate are expected. The sweep preserves prior state on failure — never wipe last-transcript.txt on a failed fetch.
• Backgrounded shell loops fight the harness. The wake signal stops arriving in chat and the user feels abandoned. Use CronCreate, always.
• The cron prompt is a string; embed the URL literally inside it so each fire is self-contained. Do not rely on session env vars surviving between fires.