Tactiq

Tactiq Notes Sync (Tactiq → projects//docs/meetings/)

On-demand sync of Tactiq.io meeting transcripts into per-project meeting archives.

Tactiq is a Chrome extension that auto-records Google Meet sessions and saves transcripts to a Google Drive folder named Tactiq Transcription/ as Google Docs.

The skill writes processed-file IDs to .tactiq_processed.json and saves transcripts under projects/<slug>/docs/meetings/.

What this skill does

1. Lists files in gdrive:"Tactiq Transcription/" via rclone.
2. For each unprocessed file, exports the Google Doc as plain text via the Drive API.
3. Parses title, date, attendees, and transcript body from the Tactiq format.
4. Routes the meeting to a project via the ## Meetings block in each project's CLAUDE.md.
5. Saves to projects/<slug>/docs/meetings/YYYY-MM-DD-<title-slug>.md.
6. Records the Drive file ID in .tactiq_processed.json (workspace root) for idempotent re-runs.
7. Anything unmatched goes to inbox/meetings/ for manual triage.

Inputs

The user invokes the skill with no required args. Optional:

• A date range (e.g., /tactiq since 2026-04-01 or last 7 days). Default: all unprocessed.
• A specific project slug to limit routing (e.g., /tactiq procure).

Prerequisites

• rclone configured with a Drive remote — gdrive on the host, gdrive-ro in sandboxed sessions (see ~/.config/rclone/rclone.conf). The skill only reads from Drive, so the read-only remote is sufficient; use whichever exists for rclone commands. The helper script finds the remote on its own.
• Helper script: research-kit/skills/tactiq/tactiq_helpers.py (Drive download, Tactiq parser, meeting-file renderer).

Step 1 — list Tactiq docs

rclone lsjson gdrive:"Tactiq Transcription/"   # sandboxed sessions: gdrive-ro:

Returns JSON with one entry per doc, including ID (Drive file ID) and ModTime. Tactiq names every file Meeting Transcription.docx regardless of meeting title — the actual title is inside the doc. Always dedup by Drive ID, never by filename.

If the call fails with RATE_LIMIT_EXCEEDED, wait ~60 seconds and retry. Drive's quota is per-minute; rclone's app credentials hit it easily.

If the user passed a date filter, prune entries whose ModTime is before the cutoff.

Step 2 — load dedup state

Read .tactiq_processed.json from the workspace root:

{
  "processed": {
    "<drive_file_id>": {
      "saved_to": "projects/<slug>/docs/meetings/<filename>.md",
      "title": "<meeting title>",
      "date": "YYYY-MM-DD",
      "processed_at": "ISO8601"
    }
  },
  "ignored": {
    "<drive_file_id>": {
      "title": "<meeting title or filename>",
      "date": "YYYY-MM-DD",
      "reason": "<short reason — e.g., 'solo recording', 'not project-relevant'>",
      "ignored_at": "ISO8601"
    }
  }
}

Skip any file ID present in either processed or ignored. If the file is missing or corrupt, treat as {"processed": {}, "ignored": {}} and warn the user.

Also reconcile against already-saved files. The cache lives at the workspace root and is per-machine and untracked — a fresh checkout (new laptop, a server) starts with no cache even though the meeting files are already committed. So before processing, also scan the committed files for the Drive IDs they already carry in frontmatter:

python3 "$SKILL_DIR/tactiq_helpers.py" scan-ids .   # {tactiq_id: path} for every saved meeting

Treat a Drive file as already processed if its ID is in the cache OR in this scan. This is what prevents duplicate -2 files when the cache is absent. Optionally backfill any scan-only IDs into .tactiq_processed.json so later runs on this machine are fast.

The ignored section is for transcripts the user has explicitly marked as not worth saving anywhere (solo test recordings, unrelated calls, etc.). Add entries when the user invokes /tactiq ignore <file_id> or asks to mark specific transcripts as ignored. Never auto-populate ignored — only the user decides what to ignore.

Step 3 — fetch each unprocessed doc

rclone alone cannot download two files that share a name in the same Drive folder (its source-side dedup keeps only one), so we use the Drive API export-by-ID directly. The tactiq_helpers.py script reads the OAuth token from rclone's config, refreshes it in memory if stale (no rclone call, no config write — works with a read-only rclone.conf), and calls the Drive API. It auto-detects the remote (gdrive, then gdrive-ro, then any drive-type section); set TACTIQ_GDRIVE_REMOTE only to override.

# Download each file by ID (SKILL_DIR = this skill's base directory).
mkdir -p /tmp/tactiq
python3 "$SKILL_DIR/tactiq_helpers.py" download <FILE_ID> /tmp/tactiq/<FILE_ID>.txt

Step 4 — parse the Tactiq export

python3 "$SKILL_DIR/tactiq_helpers.py" parse /tmp/tactiq/<FILE_ID>.txt

Returns JSON with title, date (YYYY-MM-DD), attendees (list of names), highlights, transcript (raw lines), and transcript_lines (structured {timestamp, speaker, text} objects).

The Tactiq plaintext format (observed 2026-05-01):

Transcript delivered by Tactiq.io - get it for your Google Meet today!
View the full transcript ...


<DD Month YYYY> | <Title>
Attendees: <name1>, <name2>


Highlights
<highlights or boilerplate>


Transcript
<MM:SS or HH:MM> <Speaker>: <text>
...

Notes on the format:

• Tactiq uses the Google Meet event title; if the meeting had no title, it falls back to the literal string Meeting Transcription. Flag those for the user — routing on a generic title is fragile.
• Attendees are display names from Google Meet, not email addresses.
• Long meetings are sometimes truncated by Tactiq. If transcript ends abruptly mid-sentence, note it in the report.

Step 5 — route to a project

Iterate over projects/*/CLAUDE.md and look for a ## Meetings block:

## Meetings

```yaml
match:
  - title_contains: ["procurement", "fiscal enforcement"]
  - attendee_name: ["Andrei Leite", "Lucas Mariani"]
```

Matching rules:

• title_contains — case-insensitive substring against the meeting title.
• attendee_name — case-insensitive match against any attendee display name. (Tactiq does not provide emails, so email-based attendee: rules will never match a Tactiq export.)
• A project matches if any rule matches.
• If multiple projects match, ask the user.
• If the user passed project slugs (e.g., /tactiq procure), restrict routing to those.
• If no project matches, save to inbox/meetings/ and surface this in the summary so the user can update routing rules.

Step 6 — save the file

Decide the destination path, then let the helper render and write it. Do not hand-write the markdown — the save subcommand is the single source of truth for the file format, so the layout never drifts between runs.

Destination path: projects/<slug>/docs/meetings/YYYY-MM-DD-<title-slug>.md (unmatched meetings go to inbox/meetings/YYYY-MM-DD-<title-slug>.md).

Slug rules: lowercase, ASCII-fold, non-alphanumeric → -, collapse repeats, trim, truncate to ~50 chars.

python3 "$SKILL_DIR/tactiq_helpers.py" save \
  /tmp/tactiq/<FILE_ID>.txt <FILE_ID> \
  projects/<slug>/docs/meetings/YYYY-MM-DD-<title-slug>.md

The subcommand parses the transcript, renders the frontmatter + transcript markdown, appends -2, -3, ... if the path already exists, and prints the final path written. Record that printed path in .tactiq_processed.json.

Pass --attendees "Name One,Name Two" to override the frontmatter attendee list when the Tactiq Attendees: line is ambiguous — e.g. a single Lastname, First name that the comma-split parser reads as two people.

The rendered file looks like:

---
title: <meeting title>
date: YYYY-MM-DD
attendees: [<name1>, <name2>, ...]
tactiq_id: <drive_file_id>
source: tactiq
---

# <meeting title>

## Highlights

<highlights section, verbatim — omitted when empty or default boilerplate>

## Transcript

[MM:SS] **<Speaker>**: <text>

[MM:SS] **<Speaker>**: <text>

...

The body is kept raw — the helper never summarizes, restructures, translates, or strips content.

Step 7 — update dedup state and report

After each successful save, append the entry to .tactiq_processed.json and write back. If the save fails, do not record it.

Report a summary to the user:

• N transcripts processed
• Per project: count + list of new files (relative paths)
• Unmatched: count + list of files in inbox
• Skipped (already processed): count
• Any meetings whose title is the literal Meeting Transcription — flag so the user can rename in Drive or set a Meet event title next time
• Any transcripts that look truncated

For unmatched items, suggest adding a routing rule to the relevant project's CLAUDE.md.

Bootstrapping a project's routing block

If the user asks to "set up tactiq routing for procure" or similar, add a ## Meetings block to that project's CLAUDE.md with placeholder title_contains and attendee_name rules they can edit. Don't add the block proactively — only on request.

Gotchas

• Tactiq names every doc Meeting Transcription.docx — never dedup or address by filename, always by Drive file ID.
• The first run after a long gap may pull many transcripts. Confirm with the user before processing more than ~10.
• The Drive access token in rclone.conf is short-lived (~1 hour). The helper refreshes it in memory on a 401, so no rclone pre-call is needed — but plain rclone commands (Step 1's lsjson) refresh via rclone itself; in a sandbox where rclone.conf is read-only, rclone logs a "Failed to save config" error while still working for the current command.
• Drive API quota errors (RATE_LIMIT_EXCEEDED) happen surprisingly often with rclone's app credentials. Sleep ~60s and retry.
• Tactiq exports do not include attendee emails. Email-based attendee: routing rules will never match a Tactiq export — use attendee_name: instead.
• Tactiq sometimes uses the literal title Meeting Transcription when the Meet event had no name. The slugged filename will collide if multiple such meetings happen on the same day; the -2, -3 suffix logic handles it, but flag the user so they set Meet titles up front.
• The skill does NOT delete or move docs in Drive. Files stay in Tactiq Transcription/ indefinitely.
• If .tactiq_processed.json is missing or corrupt, treat as empty and warn the user before reprocessing.
• Works in both host and sandboxed sessions; the sandbox exposes the read-only gdrive-ro remote, which is enough since the skill never writes to Drive.