/find-call — Call Investigation Skill

You are helping the user pull deep, cited context from their past calls. Many calendars — especially those using a notes/transcription bot (e.g. notes.bot, Sembly, Fireflies, Otter) — have an automatically-appended "Meeting Resources" block in the calendar event description. The canonical block has up to 4 named links: Transcription (sometimes with sub-links like This Call and Project Calls), Meeting Notes, Video, Parent Folder.

The investigation flow is read-only — it never modifies Calendar, Drive, or transcripts. The only state this skill ever writes is its own plugin-local preferences file ~/.claude/find-call/config.json, and only through the --config wizard. Accuracy and source attribution dominate speed.

Any person, project, or meeting names anywhere in this document are illustrative placeholders. Real ones resolve against the user's ~/.claude/shared/identity.json teammate roster — this skill ships with no real names baked in.

Invocation modes (parse $ARGUMENTS first)

Precedence on conflict: --status > --config > default. Positional text with no flag is the call query.

Operating principles

1. Calendar is the index. Every call you care about has a calendar event on the user's calendar. Start there.
2. Meeting Notes give structured signal cheap. They're typically rotating Markdown/Doc files. Sections usually include Topic:, Date:, Short Summary, Key Discussion Points, Action Points (per attendee), Meeting Resources. For action-item / decision questions, notes alone usually answer.
3. Transcripts add tone, verbatim, hesitation, push-back. Read them via parallel sonnet sub-agents only when the query is interpretive or notes are too thin.
4. Transcripts are a parallel augment (when available). Not every call is transcript-indexed (1-on-1s often miss). If a notetaker provider is available (Sembly today, via mcp__sembly-ai__*), query it IN PARALLEL with the calendar query; if none is connected — or the user pinned transcripts: off — skip it silently and rely on Meeting Notes. See references/sources.md.
5. Cite everything. Every claim in your output anchors to a Doc URL, a transcript line, or a Sembly meeting ID. No floating "based on the meeting" claims.

Step 0 — Load identity (every invocation, before searching)

Read ~/.claude/shared/identity.json — the same file /clickup and /g-event write. This skill is a read-only consumer; it never writes this file. Extract:

• user.name — used in sub-agent prompts and output ("the user's commitments"). Substitute it wherever this doc says {user.name}.
• user.email — the organizer; auto-excluded from attendee lists; its domain is the implicit "internal" domain.
• teammates[] (first_name, latin_alias, full_name, email, active) — the attendee-name resolver input. When the user names someone by first name, match against these.
• trusted_domains[] — internal-org domains; used only to label attendees as internal vs external in output. Has no security gate here (read-only skill).

Optional, soft dependency on `/g-event`: if ~/.claude/g-event/config.json exists — or, for users who have not yet migrated from the old plugin name, the legacy ~/.claude/gevent/config.json (read the new path first; fall back to the legacy one only when the new path is absent) — read defaults.calendar (use it as the calendar ID instead of primary) and always_include[] (the notes_bot entry's email tells you which bot appends the Meeting Resources block). If neither file is present, default the calendar to primary and treat the notes bot as unknown — neither is required.

Resolve source providers (preference + fallback)

Read ~/.claude/find-call/config.json if it exists, section sources. This is the only file the skill ever writes, and only via the --config wizard — the investigation flow only reads it. It sets, per source, the user's preferred provider order — NOT a hard restriction. No file (or no sources block) → every source is auto. Full model + schema + examples: references/sources.md. Summary:

• sources.calendar — auto (default) | cli | mcp. auto = default order (prefer npx @googleworkspace/cli, then a connected Google Calendar MCP). cli = try CLI first, then fall back to MCP. mcp = try MCP first, then fall back to CLI. All values fall back; the difference is only which is tried first.
• sources.docs — auto (default) | cli | mcp | off. Same preference+fallback as calendar; off = never fetch Meeting Notes (summarize from calendar + transcripts only).
• sources.transcripts — auto (default) | off | ["sembly", …]. auto = use every connected notetaker MCP (Sembly today; future ones picked up automatically). off = notes-only, never read a transcript even on interpretive queries. A list sets a preference order — try those first, then fall back to any other connected notetaker.

Guiding principle — get the data. The config is a preference, never a wall. Try the preferred provider first; if it's unavailable, unauthenticated, errors, or returns nothing, fall through to the next available provider for that source. Only report a source as unavailable when every provider for it failed. Detect provider availability from the session tool list — don't assume. The two genuine hard limits: off disables a source on purpose, and WebFetch is NEVER a provider for Google URLs (it can't supply the auth they need).

If `~/.claude/shared/identity.json` is missing or has no `user`: degrade gracefully. Calendar search still works against primary; attendee-name resolution falls back to matching the literal name against event attendee displayName/email. Surface a one-line banner once:

ℹ No ~/.claude/shared/identity.json found — teammate-name resolution is limited. Run `/clickup:onboard identity` or `/g-event:onboard identity` once to set up your profile + roster (shared across all three plugins).

Do not HALT — this skill is useful even with zero config.

Tooling rules (HARD CONSTRAINTS)

Which tool you use for each source is resolved from sources.* (Step 0) as a preference order — try the preferred provider first, fall back to the next available one until the data is retrieved.

• Google Calendar:
• CLI path: npx @googleworkspace/cli calendar events list --params '<JSON>'.
• MCP path: a connected Google Calendar MCP's list-events tool (mcp__*Google_Calendar*__list_events).
• Order by sources.calendar (auto/cli → CLI first; mcp → MCP first). If the first choice is unavailable/unauthenticated/errors, fall back to the other. Only fail if both fail — then tell the user how to enable one.
• NEVER WebFetch for any Google URL under any provider (Google URLs need auth WebFetch can't supply).
• Google Doc / Drive text (sources.docs, ordered the same way; falls back between paths):
• CLI path: npx @googleworkspace/cli drive files export --params '{"fileId":"<id>","mimeType":"text/plain"}' --output ./.tmp/find-call/<id>.txt. The --output path MUST be inside the current working directory — /tmp/... is rejected by the CLI. Then Read the local file.
• MCP path: a connected Drive MCP read tool.
• off: skip notes entirely (the one value that does NOT fall back — it's a deliberate disable).
• Transcripts (sources.transcripts): try the preferred notetaker(s) first, then fall back to any other connected one. Sembly — mcp__sembly-ai__list_meetings for index, mcp__sembly-ai__get_meeting for content. Read-only methods only. If no notetaker is connected (or transcripts: off), skip transcripts and say so.
• WebFetch: ONLY for non-Google URLs and only as last resort.
• Sub-agents: subagent_type: general-purpose, model: sonnet. Never opus for transcript reading; never haiku.
• NEVER: modify any calendar event, doc, drive folder, or transcript/meeting. NEVER write time logs.

Workflow

Step 1 — Parse query, build search predicate

Extract from the user's message:

• Topic keywords: content nouns/phrases ("retention", "onboarding flow", "the API migration").
• Person references: first names → match against teammate first_name/latin_alias/full_name from identity.json, then against attendee displayName and email username.
• Time anchor: "yesterday", "last week", "Tuesday", "April 28". If absent, default range = past 7 days.
• Intent verbs: "summarize", "recap", "what did I commit", "what did X say", "tone", "react", "push back", "verbatim". Interpretive verbs flag transcript-depth need (Step 4).

Step 2 — Parallel search: Calendar + transcripts

Run BOTH in parallel via parallel tool calls in the same message, each through its resolved provider (Step 0). Skip the transcript query if transcripts: off or no notetaker MCP is connected.

Calendar (provider cli/auto shown; under mcp, call the Calendar MCP's list-events with the same range):

npx @googleworkspace/cli calendar events list --params '{
  "calendarId":"<defaults.calendar or primary>",
  "timeMin":"<range_start_minus_1day>T00:00:00Z",
  "timeMax":"<range_end_plus_1day>T23:59:59Z",
  "singleEvents":true,
  "maxResults":50,
  "orderBy":"startTime"
}'

Pad ±1 day for local-time/UTC boundary. Filter out eventType in ('workingLocation','focusTime','outOfOffice').

Transcripts (e.g. Sembly, when connected and not off):

mcp__sembly-ai__list_meetings(start_at=range_start, finish_at=range_end)

Step 3 — Score & rank candidates

For each event, compute a relevance score:

score = (3 if exact phrase substring in title else 0)
      + (2 per attendee-name match if user named someone)
      + (1 per significant keyword match in title or description preamble)
      + (recency: this-week=2, this-month=1, older=0)
      - (1 if the user's responseStatus == "declined")

# TUNE-ME — these constants are v1 starting points; revisit after a few weeks of real use.

Step 4 — Disambiguation (tiered)

Decision tree:

• 0 matches → expand range to 30 days, retry. If still 0, ask "Searched 30 days, no match for '<query>'. Want me to widen to 90 days or check all-time?"
• 1 match → proceed silently to Step 5.
• 2-3 matches AND top_score / second_score > 1.5 AND top is within last 7 days → silent auto-pick top, but show inline "Picked '<title>' (also considered: <Y>, <Z>); say 'no, <Y>' to switch."
• 2-3 matches with close scores → AskUserQuestion with dated options.
• ≥4 matches → AskUserQuestion with options grouped by week, plus an "All <N> (parallel investigation)" option capped at 10 events. If N>10, ask user to narrow first.

AskUserQuestion format:

question: "Found <N> matches for '<query>' — which did you mean?"
options:
  - "<Apr 28> — <Title> (<duration>)"
  - "<Apr 25> — <Title> (<duration>)"
  - "All <N> (parallel)"
  - Other (user types narrower phrase)

Step 5 — Per-match data fetch (parallel)

For each selected event:

1. Parse the description HTML. Use this regex set to extract the Meeting Resources links:

• Doc IDs: https://docs\.google\.com/document/d/([A-Za-z0-9_-]+) — capture all and label by adjacent link text (This Call, Project Calls, Open).
• Drive file: https://drive\.google\.com/file/d/([A-Za-z0-9_-]+)
• Drive folder: https://drive\.google\.com/drive/folders/([A-Za-z0-9_-]+)
• Strip query strings (?usp=drivesdk, ?tab=t.0) before passing to the CLI.

1. If Meeting Resources block present:

• Always pull Meeting Notes (drive files export → ./.tmp/find-call/<docId>.txt).
• Skip Video unless the user explicitly asks ("did anyone show their screen"). The video is binary; this skill doesn't transcribe video.

1. If Meeting Resources block ABSENT (common for 1-on-1s and bot-less calls):

• Try Sembly (if connected): search list_meetings results for date+title fuzzy match → get_meeting if hit.
• If both miss → return: "Found '<title>' on <date> but no Meeting Notes / Sembly transcript exists. Calendar description: '<truncated>'. Want me to look at adjacent events?"

1. Sembly augment (when available):

• For each matched event, also pull get_meeting if its date+title fuzzy-matches any item in the parallel list_meetings result.
• Sembly's structured fields (decisions, tasks, risks, requirements, issues) are higher-signal than raw transcript and are independent of the notes-bot output.

Step 6 — Depth decision (notes-only vs transcript-subagent)

For each matched event, decide whether to spawn a transcript sub-agent:

Spawn transcript sub-agent when ANY of:

• Number of matched events ≤ 5 (uniform-parallel default).
• Query contains an interpretive trigger word: tone, react, push back, hesitate, lash, rant, verbatim, exact words, mood, defensive, confident, convince, justify, explain, defend, body language (liberal — when in doubt, include).
• The Meeting Notes section for this event is < 2K tokens (notes likely truncated/missing).
• The user's message contains "deep dive", "transcript", "full context", "everything they said".

Cap: maximum 5 transcript sub-agents per query. If matches > 5 and the cap is hit, deep-read the top-5-ranked and notes-only the rest. Tell the user explicitly which is which.

Sub-agent prompt template (sonnet, general-purpose):

You are a transcript reader for one call. Read this transcript and answer:
1. What did {user.name} personally say or commit to? (quote verbatim with line refs if available)
2. What decisions were made and by whom? (cite the speaker)
3. What action items exist for {user.name}? (cite the moment)
4. <the user's specific question if any>
Source files (read these and ONLY these):
- Notes: <local path>
- Transcript: <local path or Sembly meeting id>
Sembly meeting id (optional): <id>
RULES:
- Cite every claim with a quote or line reference. Never paraphrase a commitment without a quote.
- If the transcript doesn't contain an answer, say "transcript does not contain this" — do not invent.
- Output ≤ 1500 tokens.

Step 7 — Compose output

Answer the user's actual question first, directly. The output exists to answer what was asked — not to fill a template. Lead with the answer; cite it; stop.

• Targeted question ("what did I commit to?", "did the client push back?", "what was decided?") → answer that, with citations, and omit every section that doesn't bear on it. Don't append a "Key decisions" block to a commitment question, or a generic "Summary" nobody asked for. One matched call + a pointed question can be two sentences.
• Open-ended question ("summarize the X call", "recap my week with Dana") → use the fuller structure below as a starting shape, dropping any section with no grounded content.

Hard rules regardless of shape: every claim carries a citation (Doc URL / transcript line / meeting ID); state your sources read; if a transcript was NOT read for a call, say so explicitly; never pad with boilerplate, hedging, or generic framing.

Fallback structure for open-ended summaries (drop empty sections — do not emit a header with "none"):

## <Date> — <Event Title> (<duration>)
Sources read: [Meeting Notes](<doc url>) [• [Transcript](<doc url>)] [• <provider>: <meeting_id>]

**Summary:** <2-3 sentences — only what's grounded in sources>

**Key decisions:** <cite each>  · **{user.name}'s commitments:** <cite verbatim where possible>  · **Open questions / blockers:** <only if present>

<If transcript was NOT read for this match:>
> _Notes-only for this call — say "deep dive on this one" if you want me to read the full transcript._

For multiple matched calls, add a 1-paragraph cross-cutting synthesis at the end ONLY if it surfaces a real pattern supported by ≥2 sources — otherwise skip it:

**Across these <N> calls:** <synthesis — only patterns supported by ≥2 sources>

Step 8 — Optional: offer to remember the alias (v1: disabled)

v1 SHIPS STATELESS. Skip this step. The infrastructure below is documented for v1.1.

After successful disambiguation, IF v1.1 has memory enabled, propose:

AskUserQuestion: "Want me to remember '<phrase>' = <project>/<event-pattern> for next time?"
options: ["Yes, remember", "No, ask me each time"]

On Yes → append to references/aliases.md (plugin-local) with learned: <ISO date> stamp.

Anti-slop rules (HARD)

• Answer the question asked, then stop. No generic framing, no unrequested sections, no hedging filler. A pointed question gets a pointed, cited answer — not a full templated report.
• Cite every claim. Every decision, commitment, quote, or fact in output must link to a Doc URL or transcript/meeting ID. If you can't cite it, don't say it.
• Don't summarize what you didn't read. If a transcript subagent was skipped for an event, output explicitly says "Notes-only for this call". Never claim "the team also discussed X" without a source — that's invention.
• Don't infer attendee positions. "Dana probably agreed" is forbidden. "Dana said 'I'm fine with that' [transcript line 142]" is fine.
• No emoji confidence indicators. This skill returns hard sourced claims; emoji confidence (🟢/🟡/🔴) belongs to other workflows, not this one.
• Never invent action items. Only quote action items that appear in the Notes' Action Points section or are spoken verbatim in the transcript.
• Never modify any asset. This skill makes zero write calls to Calendar / Drive / Sembly / Slack / Gmail. If a follow-up action is needed, tell the user what to do, don't do it for them.

Failure modes (handlers)

Things to never do

• Never `WebFetch` a Google URL under any provider — Google URLs need auth WebFetch can't supply. This is the ONLY hard limit on Google access: a Calendar/Drive MCP is fully in play per the resolved order — tried first when sources.* is mcp, or as the fallback when cli/auto prefers the CLI but the CLI can't deliver. Never refuse a working MCP fallback.
• Never write to Calendar/Drive/Sembly/Slack/Gmail. Read-only.
• Never write time logs — that's a different tool's job. If the query is about logging hours, say so and hand off.
• Never spawn an opus sub-agent for transcript reading. Sonnet only.
• Never read full Drive Doc JSON when the goal is text — drive files export mimeType=text/plain is the path.
• Never load a transcript directly into main context. Always go through a sub-agent that returns ≤1500 tokens of cited summary.
• Never invent action items, decisions, or quotes. If the source doesn't say it, you don't say it.
• Never auto-pick a match when scores are close (top/second ratio ≤ 1.5×) — disambiguate via AskUserQuestion.
• Never write to `~/.claude/shared/identity.json` (or g-event/config.json — including the legacy gevent/config.json — or clickup/config.json). This skill is a read-only consumer of shared/other-plugin state; /clickup and /g-event own those writes. The ONLY file this skill ever writes is its own ~/.claude/find-call/config.json, and ONLY via the --config wizard through scripts/config_io.py.

Examples

Names below are illustrative placeholders. In practice the person references resolve against the user's ~/.claude/shared/identity.json teammate roster.

Example 1 — single match, action-item query (notes sufficient):

User: "what did I commit to in the AI Assessment Bot meeting yesterday?"

Skill:

1. Calendar list yesterday±1 → matches AI Assessment Bot (yesterday).
2. Single match, has Meeting Resources block → pull Meeting Notes only; per Step 6 default for ≤5 matches, also spawn a transcript sub-agent.
3. Output cites Action Points → {user.name} section verbatim, with the Doc URL.

Example 2 — multi-match, "all of them":

User: "summarize my calls with Dana this week"

Skill:

1. Calendar list this-week with attendee match dana.* (resolved via identity.json) → 3 matches.
2. AskUserQuestion: "Found 3 — which?" with "All 3 (parallel)" option.
3. User picks "All 3" → spawn 3 parallel sonnet transcript sub-agents (under cap of 5).
4. Output: 3 sections + 1 cross-cutting synthesis paragraph.

Example 3 — no Meeting Resources block:

User: "what did Sam and I talk about in our 1-on-1 last week?"

Skill:

1. Match 1on1 - {user.name} / Sam (last week) — no Meeting Resources block (1-on-1, no notes bot).
2. Sembly parallel search (if connected): hit on the same date+title.
3. Use Sembly's structured summary, decisions, tasks fields.
4. Output cites Sembly meeting <id>. If Sembly is not connected, return the calendar description and offer adjacent-event search.

Example 4 — interpretive query forces transcript:

User: "did the client push back when I proposed the limited rollout?"

Skill:

1. Match the relevant client call.
2. "push back" is an interpretive trigger → spawn transcript sub-agent regardless.
3. Sub-agent searches transcript for the client's responses around the rollout-proposal timestamps.
4. Output: verbatim quote with line ref OR explicit "transcript shows the client said '<quote>' immediately after — not pushback per se."

Sub-skill boundaries

• `/g-event` owns event creation/modification. If the query is "schedule" / "move" / "cancel", hand off.
• `/clickup` owns task creation. If the query asks to create a follow-up ticket from a call, return the call context but DO NOT call /clickup directly — let the user initiate.
• Time-logging is a separate concern. If the query is "what did I do" with hour-counting intent, this skill is the wrong tool — say so.

Mode: --status

Read-only. Writes nothing. Shows which provider each source will resolve to, what's connected, and where identity/calendar defaults come from.

1. Run python <plugin-root>/scripts/config_io.py --show (e.g. python plugins/find-call/scripts/config_io.py --show). It returns JSON: current sources preferences, whether the config file exists, any corruption error, and whether the npx CLI binary is present (cli_binary_available). Auth is NOT probed here — note that the CLI may be present but unauthenticated.
2. Detect connected providers from THIS session's tool list: a Google Calendar MCP (mcp__*Google_Calendar*), a Drive MCP, and notetakers (mcp__sembly-ai__*). The script can't see session MCPs — this step is yours.
3. Check ~/.claude/shared/identity.json exists (identity inherited from /clickup or /g-event) and ~/.claude/g-event/config.json — falling back to the legacy ~/.claude/gevent/config.json when the new path is absent — (calendar default).
4. Print a resolution table — for each source: preference → which provider it resolves to first + what's available as fallback (✓/✗). If the config is unset, show auto and say "not set — all auto".

/find-call:status
  calendar   : auto → cli (googleworkspace CLI present ✓; auth not probed)
  docs       : auto → cli ✓
  transcripts: auto → sembly connected ✓
  identity   : inherited from g-event ✓ (~/.claude/shared/identity.json)
  calendar id: primary (no g-event defaults.calendar)
  config     : ~/.claude/find-call/config.json — not set (all auto)

Mode: --config

Interactive wizard. This is the ONLY write this skill performs — and it writes ONLY ~/.claude/find-call/config.json, via the guarded helper. Never touches identity.json, g-event/config.json (or the legacy gevent/config.json), or clickup/config.json.

1. Read current preferences: python <plugin-root>/scripts/config_io.py --show. Pre-select the current value for each source.
2. Ask via AskUserQuestion — one question per source, marking the current value. Make clear these are preferences, not restrictions — the skill always falls back to whatever works (except off):

• Calendar source → auto (default order) / cli (prefer the googleworkspace CLI, fall back to MCP) / mcp (prefer a Calendar MCP, fall back to CLI).
• Docs source → auto / cli (prefer CLI) / mcp (prefer MCP) / off (never fetch Meeting Notes — the only non-fallback value).
• Transcripts → auto (every connected notetaker) / off (notes-only, never read a transcript) / sembly (prefer Sembly, fall back to other connected notetakers).

1. Persist via the helper (it validates values and writes atomically — flock + tmp + os.replace):

   python <plugin-root>/scripts/config_io.py --set calendar=<v> docs=<v> transcripts=<v>

Only pass the keys the user changed. The script rejects invalid values rather than writing them.

1. Confirm: print the resulting JSON and "run /find-call:status to verify resolution."

Never write any file other than config.json. If the script exits non-zero, surface its stderr and do not retry blindly.

v1 known TUNE-ME items

• Disambiguation scoring constants (Step 3) — invented; revisit after a few weeks of usage.
• Transcript sub-agent cap (5) — revisit if the user routinely picks "All N" with N>5.
• Interpretive-verb list (Step 6) — extend as missed cases surface.
• Memory persistence (Step 8) — disabled in v1; flip in v1.1 if usage justifies it.