MS Outlook + Teams Assistant (Desktop-first)

What this skill does

• Inbox nagging (Outlook Desktop): find messages from the last 7 days that likely need a reply, then send reminders until dismissed.
• Email reply drafting: produce concise drafts that match the user’s tone rules (conversational, spartan, polite; simple English; short; reduce redundancy; avoid em dashes).
• Teams tracking (optional): if Microsoft Graph is configured and permitted by tenant policy, track recent Teams chat messages that likely need a reply and nag similarly.

Safety defaults

• Do not auto-send emails or Teams messages.
• Create drafts in Outlook, or paste drafts into Telegram for approval.
• For reminders: send to Telegram by default; only send to Teams if explicitly enabled.

Setup (one time)

A) Outlook Desktop automation (recommended)

1. Ensure Outlook Desktop is installed and signed in.
2. Install the Python dependency (ask before doing this on the machine):

   • pip install pywin32

3. Create a config file:
   • Copy

     references/config.example.json

     →

     references/config.json

     and fill it.
   • IMPORTANT: Do not commit

     references/config.json

     if it contains personal IDs.

B) Teams via Graph (optional)

Only if you can create an Entra ID app registration and grant permissions.

• Copy

  references/config.example.json

  →

  references/config.json

  and fill

  teams.tenantId

  ,

  teams.clientId

  , and

  teams.scopes

  .
• Then run

  scripts/teams_scan.py

  once to complete Device Code sign-in.

See

references/teams-graph-setup.md

.

Core workflows

1) Scan and remind (Outlook)

Use

scripts/scan_outlook.py

.

1b) Scan Teams (Graph)

Use

scripts/teams_scan.py

.

Parameters:

• --days 7

  (default)

First run will print a device code sign-in message (follow it once).

Parameters:

• --days 7

  (default)

• --mode report|telegram

  (default: report)

• --max-items 200

Heuristics (editable in config):

• Within last N days
• Not from obvious broadcast sources
• Prefer threads where user is To: (not only CC) OR subject/body contains direct asks
• Prefer messages not replied by user (best-effort)

Output:

• A list of actionable items with: subject, sender, received time, why it was flagged.

Then:

• If

  --mode telegram

  , send a single concise reminder message with bullet items.

2) Dismiss / snooze an item

This skill uses a local state file to avoid nag loops.

• Dismiss: add the message’s

  internetMessageId

  (or subject+timestamp fallback) to the dismissed list.
• Snooze: store a

  snoozeUntil

  timestamp.

Use

scripts/state.py

helpers (or edit JSON directly if needed).

3) Draft an email reply (Outlook)

Use

scripts/draft_reply.py

.

4) Generate reminders (no send)

Use

scripts/scan_all.py

to update cached scan results, then

scripts/remind.py

to generate a Telegram-ready reminder message (it does not send). It applies:

• 1:1 Teams → remind when

  needsReply=true

• Group Teams → remind when

  mentionedMe=true

  AND

  needsReply=true

• Outlook → remind for flagged items

The agent should send the output to Telegram if non-empty.

Inputs:

• Either a message

  EntryID

  (preferred) or search by subject + recent window.

Behavior:

• Extract the thread (best-effort) + key metadata.
• Generate 2 drafts:
  • Short (2–5 sentences)
  • Normal (5–10 sentences)
• Apply tone rules from

  references/writing-style.md

  .

Outputs:

• Print drafts to stdout.
• Optionally create an Outlook draft reply (no sending) if

  --create-draft

  is set.

When you need more context from the user

Ask only what you cannot infer:

• Which email to reply to (subject / sender / when)
• The user’s intent (agree/decline/ask for info/confirm timeline)
• Any constraints (deadlines, attachments, names)

Keep questions minimal (max 3 at a time).