Meeting Notes Action Extractor

What this skill does

Reads a raw meeting notes document and extracts action items, identifying assignee, action verb, and due date for each one. Output is a structured task list that can be pasted into a tracker.

Inputs

• notes_path: markdown or plain text file with the meeting notes.
• Optional attendees: array of {name, handle} so handles can be resolved deterministically.
• Optional default_due_days: how many days out to assume when no date is given (default 7).

Steps

1. Read the file and split by paragraphs.
2. Detect candidate lines via patterns:
   • [ACTION], AI:, TODO:, Action item, @<name> will, <Name> to, <Name> takes, Owner: <Name>.
3. For each candidate, parse out:
   • Verb phrase (the actual action) — first verb after the assignment marker.
   • Owner — name preceding to/will or following Owner: or in the @mention form.
   • Due date — phrases like by Friday, by EOM, next sprint, Mar 12. Resolve with the meeting date as anchor.
4. Resolve owner to a handle from attendees (case-insensitive Levenshtein <= 2).
5. For lines without a clear due date, set due_date = meeting_date + default_due_days.
6. De-duplicate within the document: lines with the same (owner, verb_phrase) collapse.
7. Render a markdown checklist - [ ] @<handle> — <action> (due <YYYY-MM-DD>) and an associated CSV.
8. If a tracker target is configured (Linear/Jira/Asana), generate API-ready payloads but do not submit unless the user explicitly requests it.
9. Save to actions-<meeting-date>.md and actions-<meeting-date>.csv.

Output

A markdown checklist and a CSV with columns owner_handle, action, due_date, source_line. Stdout prints the count of items extracted and any items where owner resolution failed (those land in the report with owner: ?).

Verification

Re-read the notes and count how many candidate lines matched the patterns; report a coverage estimate (count_extracted / count_candidates). Sample three extracted items and verify the verb phrase is faithful to the source line. If more than 30% of items have unresolved owners, surface a top-of-file warning suggesting the user supply attendees. The CSV must round-trip through pandas without parse errors.

Edge cases

• Notes that use a generic "team will..." style: assign owner to the meeting's facilitator if known, otherwise leave blank with a TODO.
• Action items written as questions ("Should we change X?"): exclude unless explicitly tagged AI/TODO; questions aren't actions.
• Ambiguous owners (multiple attendees with the same first name): mark as "ambiguous" and require attendees mapping with full names.
• Non-English notes: pattern set is English-tuned; refuse to run without an explicit language override.