Reply email

Email Reply Skill

Draft email replies that sound like the user, not like AI.

Modes

Single reply (default)

/reply-email [message-id] or /reply-email + paste email text

Quick-reply shortcuts

• /reply-email confirm [message-id] — 1-line confirmation
• /reply-email schedule [message-id] — propose dates + calendar link
• /reply-email followup [message-id] — gentle follow-up

Batch mode

When user says "draft all" or during morning triage:

1. List all emails needing reply
2. Draft all replies without asking
3. Save all as Gmail Drafts
4. Show summary with time estimates

Step 1: Get the email

# Fetch the full message
gws gmail users messages get --params '{"userId":"me","id":"MESSAGE_ID","format":"full"}' --format json

Step 2: Load voice context

Read these files before drafting (every time):

1. skills/reply-email/email-voice-guide.md — real examples of how you write
2. skills/reply-email/learned-replies.md — corrections from past sessions
3. shared/communication-style.md — writing rules

Step 3: Draft the reply

Rules

• Open with the person's name (first name)
• Get to the point in 1-2 sentences
• Keep it short — if it can be 3 lines, make it 3 lines
• Include calendar link when suggesting meetings
• No pleasantries ("Hope this finds you well")
• No corporate phrases ("Feel free to", "Don't hesitate to")
• No AI words (leverage, navigate, unlock, empower, delve, foster, robust, holistic)

Final check

Read the draft out loud. Does it sound like a real person typing fast between calls? Or does it sound like AI being polite? If the latter, cut half the words.

Step 4: Save as Gmail Draft

Default: create a Gmail draft, NOT send.

# Build base64url-encoded RFC 2822 email
RAW=$(python3 -c "
import base64, email.message
msg = email.message.EmailMessage()
msg['To'] = 'RECIPIENT_EMAIL'
msg['Subject'] = 'Re: ORIGINAL_SUBJECT'
msg['In-Reply-To'] = 'ORIGINAL_MESSAGE_ID_HEADER'
msg['References'] = 'ORIGINAL_REFERENCES_HEADER'
msg.set_content('''REPLY_BODY_HERE''')
print(base64.urlsafe_b64encode(msg.as_bytes()).decode())
")

gws gmail users drafts create \
  --params '{"userId":"me"}' \
  --json "{\"message\":{\"threadId\":\"THREAD_ID\",\"raw\":\"$RAW\"}}"

Only send directly if user says "send it":

gws gmail +reply --message-id MESSAGE_ID --body "REPLY_TEXT"

Gmail label IDs

Configure these label IDs after creating them in Gmail (see INSTALL.md step 2b):

AI_READY_LABEL_ID=Label_XXX
INFO_LABEL_ID=Label_XXX
OPS_LABEL_ID=Label_XXX
HIGH_LABEL_ID=Label_XXX
BILLING_LABEL_ID=Label_XXX

To find your label IDs:

gws gmail users labels list --params '{"userId":"me"}' --format json

Step 5: Learn from the outcome

Immediate learning

When user gives feedback → incorporate, redraft, log correction.

Async learning (morning routine)

1. Fetch sent messages from last 24h
2. Cross-reference with drafts-log.md
3. Diff draft vs what was sent
4. Log corrections in learned-replies.md

What to learn from

• Words removed → too verbose or corporate
• Words added → missing natural phrasing
• Structure changes → wrong format
• Tone shifts → misjudged the relationship
• Shortened replies → wrote too much (most common mistake)

Log format for drafts-log.md

### [YYYY-MM-DD HH:MM] [Thread ID] [Recipient]
**Subject:** Re: [subject]
**Draft ID:** [gmail draft id]
**Our draft:**
> [exact text]
**Status:** pending | sent-as-is | sent-with-edits | discarded