Consult Transcriber

---
name: consult-transcriber
sheetId: "2.38"
description: >-
  Turn a consulting engagement session — a discovery call or an on-site
  shadowing session, either already captured in Fireflies or supplied as a
  recording — into a clean zh-TW transcript filed straight into the
  client's numbered [N] Drive project folder and linked on the CRM deal. A thin
  delegation wrapper: it invokes the installed training-srt-transcriber (local
  Whisper → timestamped .srt; transcribe or retime mode) and
  training-srt-optimizer (minimal-edit zh-TW cleanup) for ALL
  transcription/cleanup mechanics, adding only engagement context, filing, and
  CRM linkage. Fire eagerly on /consult-transcriber or whenever a
  CLIENT/engagement recording needs a transcript — "把這通訪談錄音轉成逐字稿",
  "處理 shadowing 錄音", "客戶錄音轉逐字稿", "整理客戶訪談錄音", "transcribe
  this client call", "process the shadowing recording", "transcript for the
  engagement". Distinct from training-srt-transcriber /
  training-srt-optimizer used directly (they produce raw/clean .srt for the
  TRAINING business and file into the training Drive folder — no engagement
  context, no CRM; use them for course/livestream material) and from
  consult-session-notes (downstream — structures the TRANSCRIPT this skill
  produces into a session summary + pain-point ledger).
category: sales-consultant
project: consult-transcriber
platform: claude
status: Done
author: Peter Tu
input: "A Fireflies meeting (id/URL/name — already transcribed, no upload) or a recording of an engagement session, + the engagement (deal URL / company) and the phase."
process: "Resolve engagement + phase → check base skills installed → /training-srt-transcriber for speech-to-text → /training-srt-optimizer for zh-TW cleanup (skip its training upload) → rename + file into the [N] folder → CRM note → hand off to consult-session-notes"
output: "A clean zh-TW transcript (.srt, optional Doc rendering) filed in the engagement's [N] folder (shadowing subfolder when applicable) and linked on the CRM deal"
synergy:
  - "training-srt-transcriber"
  - "training-srt-optimizer"
  - "consult-session-notes"
  - "consult-shadowing-scheduler"
house-style: exempt — verbatim capture — fidelity to the source outranks house voice

---

# Consult Transcriber

```bash
npx skills add https://github.com/peter-tu-zynkr/zynkr-skill-builder --skill consult-transcriber
```

Every engagement session — a discovery call, an on-site shadowing morning —
produces a recording, and the recording is worthless until it's a transcript in
the right place. This skill takes one recording and delivers a clean zh-TW
transcript **filed into the engagement's numbered `[N]` Drive folder** (the
shadowing subfolder when it's a shadowing session) **and linked on the CRM
deal**, so the downstream consult skills can find it without archaeology.

It is deliberately **thin**: the transcription and cleanup mechanics live
entirely in two installed base skills — `training-srt-transcriber` (local
Whisper → timestamped `.srt`) and `training-srt-optimizer` (minimal-edit zh-TW
cleanup). This wrapper only adds what those skills don't know: which
engagement, which phase, which folder, and the CRM breadcrumb.

## How this differs from its neighbours

- **training-srt-transcriber / training-srt-optimizer** — the base skills this
  wrapper delegates to. Used directly, they serve the TRAINING business: raw or
  cleaned `.srt` filed into the training Drive folder, no engagement context,
  no CRM. Use them directly for course / livestream material; use THIS skill
  when the recording belongs to a client engagement.
- **consult-session-notes** — downstream: structures the TRANSCRIPT this skill
  produces into a session summary + pain-point ledger. This skill ends where
  that one begins.
- **consult-shadowing-scheduler** — upstream for shadowing: it books the
  session and creates the `Shadowing — YYYY-MM-DD` subfolder this skill files
  shadowing transcripts into.

## Fixed facts (don't re-derive these)

- **Google account** for all Gmail/Drive/Docs tools: `peter_tu@zynkr.ai`
- **Drive parent folder** (`[2.2] 業務與顧問部門：專案`, home of the numbered `[N]` folders): `1hkXPX7OXPFOU0BcloPbJSFp8O0zArM8t` — orientation only; this skill never creates anything directly in it
- **CRM deal URL** for the report/backlink: `https://platform.zynkr.ai/deals/{deal_id}`

## Hard rules

1. **Delegate, never duplicate.** All Whisper/cleanup mechanics run by INVOKING
   the installed base skills by name. Never reach into their files via relative
   parent paths (a standalone install has no siblings), never re-implement
   their steps here. If a base skill is missing, STOP (step 2) — don't improvise.
2. **Never create the `[N]` folder.** No folder → STOP and route to
   /consult-intake or /consult-project-specialist. The ONE folder this skill
   may create is a genuinely-absent `Shadowing — YYYY-MM-DD` subfolder (step 1).
3. **Override the optimizer's delivery.** training-srt-optimizer's own final
   step uploads to the TRAINING business's Drive folder — that upload is
   skipped; this skill files the output itself (step 5).
4. **Client-facing email is ALWAYS a Gmail draft** — if Peter asks to send the
   transcript to the client, `mcp__google-workspace__draft_gmail_message` only.
   Never send.

---

## Workflow

### 1 · Resolve the engagement, the phase, and the target folder

**Deal** — from a `…/deals/{id}` URL or a company name. Preferred:
`mcp__zynkr__get_deal` / `mcp__zynkr__list_deals`. 
**Folder** — the deal's `notes` carry a `專案資料夾：<url>` backlink (written by
consult-intake / consult-project-specialist); extract the `[N]` folder id. If
the line is missing, list the parent (`mcp__google-workspace__list_drive_items`,
folder_id `1hkXPX7OXPFOU0BcloPbJSFp8O0zArM8t`) and match `[N] Company（…）` by
name. **No folder at all → STOP** and point at /consult-intake (inbound lead)
or /consult-project-specialist (meeting debrief) — hard rule 2.

**Phase** — discovery call (訪談) or shadowing? Infer from what Peter said and
the recording's filename/date; ask one question if genuinely ambiguous.

- **Shadowing** → the target is the `[N]` folder's **`Shadowing — YYYY-MM-DD`**
  subfolder (created by consult-shadowing-scheduler when the session was
  booked). Find it via `list_drive_items` on the `[N]` folder, matching the
  session date. Only if it's genuinely absent (e.g. the session was booked by
  hand), create it — `mcp__google-workspace__create_drive_file` with
  `mime_type="application/vnd.google-apps.folder"` inside the `[N]` folder.
  That is the one folder this skill may create.
- **Discovery call** → the target is the `[N]` folder root.

### 1.5 · Decide the source mode — Fireflies or a recording

**Ask this before anything else: was Fred in the meeting?** Anything Fireflies
captured is already transcribed, speaker-labelled and timestamped, so it needs no
recording, no upload and no ASR pass at all.

Set **SOURCE_MODE**:

- **`fireflies`** — Peter names the session ("逐字稿：夾子園那場"), gives a meeting
  id, or an `app.fireflies.ai/view/<id>` URL. Resolve with
  `mcp__fireflies__fireflies_search(query="keyword:\"夾子園\" from:2026-09-01")`
  and confirm the row by title + date + attendees. If Peter hands you a recording
  but the same session is also in Fireflies, **prefer Fireflies** and say so.
- **`recording`** — an audio/video file with no Fireflies counterpart (an
  in-person session, a phone call, anything Fred did not join). This is the
  original path: §2 → §3 → §4.

⚠️ **Fireflies request budget.** The free plan allows **50 API requests per day**
and every call spends one. Resolve in a single search, fetch once, never loop.

**When `SOURCE_MODE = fireflies`: SKIP §2, §3 and §4 entirely** — there is no ASR
to run and no cleanup to delegate. Instead:

1. `mcp__fireflies__fireflies_get_transcript(transcriptId="<id>")` → speaker-labelled,
   timestamped sentences (`[12:43 - 14:00] 高May: …`).
2. Convert those spans to `.srt` cues — one cue per sentence, `HH:MM:SS,mmm`
   start/end from the span, text as `<speaker>：<sentence>`. The result is the
   **clean** `.srt` §5 files; there is no separate raw artifact, so say so in §7
   rather than inventing one.
3. Carry the Fireflies permalink (`https://app.fireflies.ai/view/<id>`) into the
   §6 CRM note alongside the Drive link — it is the closest thing to the audio,
   which the free plan does not expose (`audio_url` / `video_url` come back empty).

Then continue at §5.

### 2 · PRECONDITION — are both base skills installed?

**Only when `SOURCE_MODE = recording`.** Check that `/training-srt-transcriber`
and `/training-srt-optimizer` are
available as installed skills in this environment. If either is missing,
**STOP** and print its install snippet for Peter to run — the same one-line
install command as this skill's own snippet at the top of this file, with the
`--skill` argument swapped to `training-srt-transcriber` or
`training-srt-optimizer`. Never inline the missing skill's mechanics instead
(hard rule 1).

### 3 · Delegate transcription → raw .srt

Run **/training-srt-transcriber** on the recording:

- **transcribe mode** (default) — fresh local-Whisper speech-to-text with
  audio-accurate timestamps.
- **retime mode** — only when a wording-correct transcript already exists
  (e.g. the client supplied minutes, or a prior STT pass was hand-corrected)
  and just needs timestamps re-derived from the audio.

Everything about venvs, models, devices, and timing validation is that skill's
business — follow its steps as written, including its own validate pass. The
deliverable of this step is its raw `.srt`.

### 4 · Delegate cleanup → clean zh-TW .srt (delivery overridden)

Run **/training-srt-optimizer** on the raw `.srt` for the minimal-edit zh-TW
pass (filler removal, obvious STT fixes, cue timing preserved) — its rules and
worksheet flow apply as written, **except its final Delivery step: skip the
upload to its default Drive folder** (that folder belongs to the training
business, not this engagement). Keep the validated `.optimized.srt` local; this
skill files it in step 5 — hard rule 3.

### 5 · File the transcript into the target folder

Rename the validated file to:

```
[N] {{COMPANY}}_{{PHASE}}_逐字稿_{{YYYY-MM-DD}}.srt      # PHASE = 訪談 | shadowing
```

(`[N]` = the engagement's folder number; date = the session date, not today.)
Then upload into the step-1 target folder:

```
mcp__google-workspace__create_drive_file(
  user_google_email = "peter_tu@zynkr.ai",
  file_name = "[4] 宏宇精密_shadowing_逐字稿_2026-08-12.srt",
  folder_id = "<target folder id from step 1>",
  mime_type = "application/x-subrip",
  content   = "<the validated .srt text>"
)
```

Then **offer, don't auto**: "要不要順便產一份 Google Doc 版逐字稿（客戶端好讀）？"
On yes, create a Doc from the transcript text (cue text in order, timestamps as
light section markers) via `mcp__google-workspace__create_doc`, then move it
into the same target folder with `mcp__google-workspace__update_drive_file`
(`add_parents = "<target folder id>"`) — creating directly in a folder via
`create_drive_file` returns HTTP 400 for Docs.

### 6 · CRM note — link the 逐字稿 on the deal

Preferred: `mcp__zynkr__update_deal`, appending to the existing notes (read
first, append, write back — never overwrite):

```
逐字稿（{{PHASE}} {{YYYY-MM-DD}}）：<transcript file url>
```

`mcp__zynkr__update_deal` REPLACES `notes` wholesale, so append in three steps:

1. `mcp__zynkr__get_deal(id="<deal_id>")` — read the current `notes`
2. build the new value: the existing notes, then a blank line, then the block below
3. `mcp__zynkr__update_deal(id="<deal_id>", notes="<combined>", confirm=true)`

Call it once without `confirm` to preview, then again with `confirm=true`. Never
send `notes` without the existing text in front of it — the field is overwritten,
not appended, and skipping the read loses every earlier backlink.

Escape single quotes by doubling them (`O'Brien` → `O''Brien`).

### 7 · Report + handoff

```
逐字稿已歸檔：宏宇精密 — shadowing 2026-08-12

| 產出 | 內容 |
|------|------|
| 逐字稿 | [4] 宏宇精密_shadowing_逐字稿_2026-08-12.srt（<file url>）|
| 位置 | [4] 宏宇精密（報價流程自動化）/ Shadowing — 2026-08-12 |
| Doc 版 | 未產出（step 5 詢問，Peter 未要求）|
| CRM | <deal url> — notes 已附逐字稿連結 |
| 下一步 | /consult-session-notes — 整理成 session summary + pain-point ledger |
```

Always end by pointing at **/consult-session-notes** — the transcript is an
intermediate artifact; the session summary + pain-point ledger is what the
engagement actually runs on.

---

## Why it's built this way

- **A wrapper, not a fork.** Whisper bootstrapping and zh-TW cue-editing
  discipline already live in the training pair and evolve there. Copying their
  mechanics here would mean two drifting copies; delegating by invocation means
  this skill inherits every base-skill fix for free.
- **The delivery override is the whole point of wrapping.** The base pair is
  tuned for the training business, ending with an upload to the training Drive
  folder. An engagement transcript in that folder is effectively lost. The one
  behavior this wrapper changes is WHERE the output lands.
- **Filing goes to the seam the suite already uses.** consult-shadowing-scheduler
  creates `Shadowing — YYYY-MM-DD` precisely so the session's transcript has a
  known home; consult-session-notes and consult-brd-writer read from there.
  Discovery calls land in the `[N]` root because they predate any subfolder.
- **STOP on a missing base skill.** Silently re-implementing transcription
  when the base skill is absent would fork the mechanics and hide the install
  gap. A hard stop with the install snippet is cheaper than a divergent copy.

## Inference defaults (Peter overrides by just saying so)

- **Mode** → transcribe; retime only when a wording-correct transcript is
  supplied alongside the audio.
- **Phase** → shadowing if the `[N]` folder has a `Shadowing — YYYY-MM-DD`
  subfolder matching the recording's date; otherwise ask.
- **Language** → zh (the base transcriber's default); pass through Peter's
  override for other languages.
- **Doc rendering** → OFF; offered in step 5, produced only on a yes.
- **Filename date** → the session date (from the subfolder name or the
  recording metadata), not the run date.

## Provenance

Wraps `training-srt-transcriber` (4.11) and `training-srt-optimizer` (4.10) by
delegation @ b6bfb04c (2026-08-03) — no mechanics copied; breaks only if the
base skills' invocation contracts change.

## Reference files

- None ship with this skill — a delegation wrapper carries no mechanics. The
  transcription notes and zh-TW subtitle rules live in the two base skills'
  own `references/` folders and are consulted by those skills when invoked.

## Limitations

- Requires both base skills installed — it stops (by design) rather than
  transcribe on its own; transcription quality, speed, and language support
  are exactly the base pair's.
- Requires a deal whose notes carry the 專案資料夾 backlink; it never creates
  the `[N]` folder (only a missing `Shadowing — YYYY-MM-DD` subfolder).
- One recording per run — a multi-session batch is multiple runs.
- It files and links the transcript but does not interpret it: no summary, no
  pain points, no requirements (that's /consult-session-notes onward).