Minutes Tag Agent Skill

/minutes-tag

Lightweight outcome tagging — adds an outcome: field to a meeting's frontmatter so /minutes-mirror can correlate the user's behavior with their results over time.

The whole point of this skill is speed. Tagging should take 5 seconds, not 5 questions. Don't be precious about it — most users will never adopt tagging if it feels like data entry.

How it works

Phase 1: Identify the meeting

Three patterns the user might use. Always filter to meetings, not voice memos — voice memos can't be "won" or "lost".

1. Most recent ("tag this meeting", "mark that as a win", "tag the call I just finished"):

minutes list --content-type meeting --limit 1

Use the most recent. Don't ask which one — that defeats the speed promise. The default behavior should always be "the call you just had".

2. By date ("tag yesterday's call", "tag the Tuesday call"):

minutes list --content-type meeting --limit 10

Pick the meeting matching the date. If multiple meetings match the same day, ask once: "You had <N> meetings <date>. Which one?" with options listing titles.

3. By name ("tag my call with Sarah as a win"):

minutes search "<name>" --content-type meeting --limit 5

Pick the most recent. If ambiguous, ask once.

Phase 2: Identify the tag

If the user already named the outcome in their message ("tag that as a win"), use it directly. Don't ask again — they already told you.

If they haven't, ask via AskUserQuestion with these standard options:

• won — got the outcome you wanted (deal closed, decision made, agreement reached)
• lost — didn't get what you wanted (deal lost, idea rejected, no decision)
• stalled — neither — went sideways, no clear outcome, needs another meeting
• great — high-quality conversation regardless of outcome (insight, real connection, energy, learned something)
• noise — should have been an email; no value; time wasted
• (custom) — let the user provide their own tag

Standard tags are the only ones /minutes-mirror will correlate. Custom tags are stored faithfully but won't appear in correlation analysis — warn the user gently if they pick a custom one: "Custom tags are saved, but mirror only correlates the standard five."

Phase 3: Capture a note only if the user gave one in their message

Do not ask an interactive note question. That's a second prompt and it breaks the speed promise.

Parse the user's original message for a "why" or note. Common patterns:

• "tag as won, note: Sarah committed to monthly billing" → note = "Sarah committed to monthly billing"
• "tag won — got the verbal commit on pricing" → note = "got the verbal commit on pricing"
• "tag stalled because Alex postponed the decision" → note = "Alex postponed the decision"

If you find a note in the message, use it. If you don't, leave outcome_note out of the frontmatter entirely. Don't insert an empty field. Don't ask. Users who want a fuller record have /minutes-debrief for that.

Phase 4: Edit the frontmatter via the bundled helper script

Use the script — do not Edit the frontmatter manually. YAML frontmatter is fragile, and the script handles all the edge cases (no existing frontmatter, existing outcome that needs replacement, atomic write to prevent half-edits, preservation of all other fields).

python3 "${CLAUDE_PLUGIN_ROOT}/skills/minutes-tag/scripts/tag_apply.py" \
  "<absolute-path-to-meeting-file>" \
  --outcome <won|lost|stalled|great|noise|custom> \
  [--note "the optional one-line note from Phase 3"]

Pass --note only if Phase 3 found a note in the user's message. Skip the flag entirely otherwise — the script will omit outcome_note from the frontmatter rather than inserting an empty field.

What the script guarantees:

• The new fields (outcome, outcome_note if a note was passed, tagged_at) are inserted just before the closing --- of the frontmatter, after every other existing field.
• All other frontmatter fields are preserved byte-for-byte — no reordering, no reformatting, no whitespace changes.
• Re-tagging is fully idempotent: if outcome: already exists, the script removes the old outcome lines and re-inserts fresh ones at the end. Old outcome_note: is dropped if no new note is passed.
• The body of the meeting file is never touched — only the frontmatter block.
• Writes are atomic (temp file + rename) so an interrupted run can never leave a half-written meeting.

The script prints {"status": "ok", ...} to stdout on success, or {"error": "..."} to stderr with non-zero exit on failure. Surface any error to the user.

Fallback if Python isn't available (extremely rare on macOS): use the Edit tool with surgical precision. Find the closing --- of the frontmatter, anchor on a small unique block ending in it, and insert your new fields right before. This is brittle on unusual frontmatter — only do it if the script fails.

Phase 5: Confirm and nudge

Confirm in one line: "Tagged <meeting title> as <outcome>."

Then verify the file is still parseable by Minutes after the edit. The slug is the filename minus .md (e.g., 2026-03-18-product-roadmap-with-case):

minutes get "<filename-without-.md>" 2>&1 | head -3

If the output contains an error or warning about malformed frontmatter, surface it gently: "Note: this meeting's frontmatter has a pre-existing schema issue. The tag was saved, but /minutes-mirror may skip this meeting until it's fixed." Don't try to fix the unrelated schema issue — that's not tag's job.

One-time lifetime nudge (idempotent — never repeats):

ls ~/.minutes/tag-nudge-shown 2>/dev/null

If that marker file doesn't exist, this is the first time tag has run on this machine. Show the nudge once, then create the marker:

> "First tag — nice. When you've tagged ~10 meetings, run /minutes-mirror trends and I'll show you what your winning meetings have in common."

mkdir -p ~/.minutes && touch ~/.minutes/tag-nudge-shown

The marker file is the state. No counting, no edge cases, no risk of repeated nudges from re-tagging the same meeting.

Gotchas

• Speed is the entire feature. If tagging takes more than two questions (the tag, optionally the note), you've broken it. Default to "most recent". Skip the optional note unless the user clearly wants to add one.
• Standard tags only correlate. Mirror's correlation analysis only works on the five standard tags: won, lost, stalled, great, noise. Custom tags are saved but won't be analyzed. Warn the user once if they pick a custom tag — don't lecture them, just let them know.
• Don't touch the meeting body. Only edit the YAML frontmatter block between the first two --- markers. Use Edit with surgical precision.
• Re-tagging is intentional. If the user tags a meeting that's already tagged, overwrite it cleanly. They're either correcting themselves or seeing it differently after the fact. Both are valid.
• Preserve existing frontmatter exactly. Some meetings have action_items, decisions, intents, entities, people, calendar_event, captured_at, device, recorded_by, etc. Don't reformat or reorder anything — only insert/update the three outcome fields.
• Tag freshness matters. Tags are most valuable within ~24 hours, while the outcome is fresh in the user's head. Tagging two weeks later is fine but worth less. Don't enforce this — just don't make tagging feel like a chore that the user puts off.
• Don't try to infer the tag from the transcript. If the user says "tag this meeting" without saying which outcome, ask. Don't guess from the transcript — your guess will be wrong in the cases that matter most (a meeting that looks like a win on paper but actually wasn't, or vice versa).
• The note is optional for a reason. Most users will skip it. That's fine — the tag itself is the load-bearing data. Don't make the user feel like they're underperforming if they skip the note.