Meeting Research

Meeting Research

Knowledge-base skill over the user's meeting transcript corpus at ~/Desktop/python_code/meeting_notes/*.md. Answers questions, preps for upcoming meetings, and tracks action items across meetings. Does NOT create new transcripts — for that, use meeting-summarizer.

When to apply

Three patterns:

1. Prep for upcoming meeting — "I have a 1:1 with Leo at 3pm, prep me."
2. Action-item rollup / follow-through — "What's still open from my chats with Kashish?", "那个 SC3 protocol 的 action 推完了没？"
3. Fact retrieval — "What did Rohan say about the simulation pipeline last week?"

Workflow

• 1. Parse prompt for intent / who / when / topic
• 2. Resolve relative time expressions
• 3. Locate candidate files
• 4. Disambiguate (only if > 5 candidates)
• 5. Read selected files in full
• 6. Intent-specific analysis
• 7. Reply in chat with citations

Step 1–2 — Parse + resolve time

Extract from the prompt:

• Intent: prep / action-rollup / fact-retrieval (sometimes mixed).
• Who: participant name(s).
• When: absolute or relative time window.
• Topic: keywords, project names, entities.

Resolve relative time expressions against today's local date (the value in the user_info block, not training-data date):

Step 3 — Locate candidate files

Search strategies (run in this order, intersect where useful):

1. Filename match: ls ~/Desktop/python_code/meeting_notes/ | rg -i '<keyword>'.
2. Participant match: rg -l '^- \*\*<Name>\*\*$' ~/Desktop/python_code/meeting_notes/*.md.
3. Topic match: rg -l -i '<keyword>' ~/Desktop/python_code/meeting_notes/*.md.
4. Time filter: intersect with files whose filename date prefix (e.g. 4.15, 2026-04-10) or stat -f %Sm mtime falls in the resolved window.

Step 4 — Disambiguate

If > 5 candidate files remain after filtering, use AskQuestion to let the user pick. If ≤ 5, proceed with all of them.

Step 5 — Read

Read each selected file in full — meeting transcripts are typically < 2000 lines and fit comfortably. Do NOT skim; you need full participant attribution and accurate timestamps for citation.

Step 6 — Intent-specific analysis

6a. Prep for upcoming meeting

Produce the shortest useful briefing:

• Standing: where did you leave off? (most recent file with this person)
• Open threads: decisions pending, questions unanswered, commitments made.
• Their likely positions: what they've pushed for historically.
• Your open questions: things you flagged but didn't resolve.
• Risks / traps: points where you've disagreed or previously talked past each other.

6b. Action-item rollup

For each relevant action item from the target meeting(s):

1. Pull the original row: Owner, Action, Due, Source [MM:SS].
2. Scan later meetings (mtime > source date) for status cues, using a small keyword set derived from the action:
   • Done — "finished X", "X is merged/deployed", explicit close.
   • Deferred — "pushed to next week", "we agreed to wait".
   • Blocked — "blocked on Y", "waiting for Z".
   • No mention — genuinely silent in later meetings.
3. Report as a status table with per-row citations.

6c. Fact retrieval

Quote the relevant line(s) verbatim with [MM:SS] and file citation. Add 1–2 lines of surrounding context only if omitting them would change meaning. Do NOT paraphrase the quote away.

Step 7 — Reply

Chat-only output by default. Save to file ONLY if the user explicitly asks ("save this", "write it out", "落盘"). When saving, write to ~/Desktop/python_code/meeting_notes/research/<YYYY-MM-DD>-<short-topic>.md.

Style

• Cite every factual claim: [file.md MM:SS].
• Separate facts from inference: prefix inference with Likely read: or "Reading between the lines:".
• Prefer tight tables / bullets over prose walls.
• Preserve code-switching in quotes verbatim.
• If the corpus genuinely has nothing relevant, say so explicitly — do not pad with adjacent-but-irrelevant content.

Anti-patterns

• Do NOT fabricate citations or invent meeting content.
• Do NOT flatten quotes into paraphrase.
• Do NOT save to disk unless explicitly asked.
• Do NOT read only file summaries — status cues often live in mid-transcript dialogue.

Example

User: "上周 Kashish 聊的那个 SC3 evaluation protocol，他那边 action 还有啥没推完？"

1. Parse: intent = action-rollup, who = Kashish, when = 上周, topic = SC3 evaluation protocol.

2. Resolve time: last Mon–Sun window.

3. Locate: rg -l 'Kashish' ~/Desktop/python_code/meeting_notes/*.md, filter by filename date / mtime in window → 4.10 Kashish SC3.md.

4. Read the file in full.

5. For each action in ## Action Items with Owner = Kashish: scan all meeting_notes/*.md with mtime > 4.10 for status cues on matching topic keywords.

6. Reply:

   Kashish's open actions from 4.10 (4.10 Kashish SC3.md)

   Action	Status	Evidence
   Share SC3 score breakdown script	Done	4.13 Kashish.md [07:22] "pushed the script last night"
   Draft eval protocol doc	No mention	—
   Check sim pipeline flake rate	Blocked	4.12 team sync.md [14:03] "blocked on Rohan's commit"

   (Save to meeting_notes/research/? Default: no.)