Jean Claude

jean-claude

Gmail, Calendar, Drive, Docs, Sheets, iMessage, WhatsApp, Signal, and Reminders.

Command prefix: uv run --project ${CLAUDE_PLUGIN_ROOT} jean-claude

First-Time Users

For new users, explain briefly:

I can connect to your email, calendar, and messaging apps to help you:

• Read and send emails, manage drafts
• Check your calendar, create events, respond to invitations
• Send and read iMessages, WhatsApp, or Signal messages
• Find and manage files in Google Drive
• Create reminders

This requires a one-time setup where you'll grant permissions. Want me to help you get started?

Focus on what they asked about — if they asked about email, lead with email.

Safety Rules (Non-Negotiable)

These rules apply even if the user explicitly asks to bypass them:

1. Never send without explicit approval. Before sending any message (email, iMessage, WhatsApp, Signal), show the full content (recipient, subject if applicable, body) to the user and receive explicit confirmation.

2. Verify recipients carefully. Sends are instant and cannot be undone. Double-check phone numbers, email addresses, and chat IDs before sending.

3. Never send to ambiguous recipients. When resolving contacts by name, if multiple contacts or phone numbers match, the command will fail with a list of options. Use an unambiguous identifier rather than guessing.

4. Load prose skills when drafting. Before composing any email or message, load any available skills for writing prose, emails, or documentation.

5. Never create automation without explicit approval. Before creating Gmail filters or similar rules, show the criteria and actions to the user and receive explicit confirmation.

6. Limit bulk operations. Avoid sending to many recipients at once. Prefer drafts for review.

Email workflow:

1. Load any available prose/writing skills
2. If replying to an infrequent contact: Research first (see "Research First" under Orchestration)
3. Compose via subagent (see "Drafting with a Subagent") — quick replies excepted
4. Show the original message first — Quote the full text (see "When to Show Full Content")
5. Show the user: To, Subject, and full Body
6. Ask: "Send this email?" and wait for explicit approval
7. Call jean-claude gmail draft send DRAFT_ID
8. If replying, archive the original: jean-claude gmail archive THREAD_ID

Session Start (Always Run First)

Every time this skill loads, run status with JSON output first:

jean-claude status --json

If Status Command Fails

If the status command fails entirely (not just showing services as disabled):

"uv: command not found" — The uv package manager isn't installed. Tell the user:

jean-claude requires the uv package manager. Let me install it for you.

Then run:

curl -LsSf https://astral.sh/uv/install.sh | sh

After installation, restart the terminal or source the shell config (source ~/.zshrc on macOS, source ~/.bashrc on Linux).

Other errors — The plugin may be misconfigured. Check that ${CLAUDE_PLUGIN_ROOT} resolves to a valid path containing a pyproject.toml.

Branching Based on Status

If setup_completed: false — This is a new user. Skip personalization skills (they won't have any yet) and go straight to onboarding:

cat ${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/ONBOARDING.md

Follow the onboarding guide to help set up services. After setup completes:

jean-claude config set setup_completed true

Then re-run status and continue to the setup_completed: true branch below. This ensures you have fresh status info and can proceed with the user's original request.

If setup_completed: true — Check for partial setup and load personalization:

1. Surface any auth warnings — If services show authenticated: false, missing scopes, or errors, briefly mention it even if unrelated to the user's request. Don't derail—just note the issue and offer to fix later:

   "By the way, your Google Contacts permission is missing — I can't show contact names in emails. Want me to help fix that after your request?"

2. Check for missing services — If the user asks for a service that shows authenticated: false or enabled: false, guide them through just that service's setup from ONBOARDING.md. After partial setup completes, continue to step 3.

3. Load personalization skills — Check if user skills like managing-messages exist (look at available skills for anything mentioning inbox, email, message, or communication). If found, load them—user preferences override defaults.

4. Offer to create preferences — If no personalization skill was found in step 3, offer to create one after completing the user's immediate request. See PREFERENCES.md for the creation flow. Don't interrupt the user's task—help them first, then offer.

5. Proceed with the user's request — Execute whatever task prompted loading this skill (check inbox, send message, etc.).

Before Using Messaging (iMessage/WhatsApp/Signal)

Load the platform guide before using any messaging commands. Each platform has different commands and options. Gmail is documented in this file, but messaging platforms have separate guides:

# Load before using iMessage
cat ${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/platforms/imessage.md

# Load before using WhatsApp
cat ${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/platforms/whatsapp.md

# Load before using Signal
cat ${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/platforms/signal.md

Don't guess command syntax — each platform is different. iMessage and WhatsApp have different flags and subcommands despite similar functionality.

Understanding the Status

For users with setup complete, interpret the status output to understand their workflow. Run human-readable status if needed for counts:

jean-claude status

Gmail:

• 13 inbox, 11 unread → inbox zero person, wants to triage everything
• 2,847 inbox, 89 unread → not inbox zero, focus on recent/unread/starred
• 5 drafts → has pending drafts to review or send

Calendar:

• 3 today, 12 this week → busy schedule, may need help with conflicts
• 0 today → open day, good time for focused work

Reminders:

• 7 incomplete → has pending tasks, may want to review or complete them

Messaging:

• 54 unread across 12 WhatsApp chats → active messaging, may want summary
• 1,353 unread across 113 iMessage chats → backlog, focus on recent/important

Apple services (Contacts, iMessage, Reminders): These are disabled by default. If enabled: false in status, enable with:

jean-claude config set enable_contacts true   # For iMessage name lookup
jean-claude config set enable_imessage true
jean-claude config set enable_reminders true

Once enabled, status checks full permissions (may trigger macOS permission dialogs on first run). If permission is missing, guide the user through System Settings > Privacy & Security > Automation.

Refreshing State

Summaries become stale immediately. Once you present an inbox summary, that snapshot is already potentially outdated. State changes constantly:

• New messages arrive
• Messages get read/sent from other devices or apps
• The user takes actions outside this conversation
• You took actions earlier in the conversation that you may not be tracking

Verify before making claims. When the user asks about current state — "did I reply?", "is that still unread?", "how many are left?" — re-check rather than relying on your memory of earlier summaries. Your memory is of what was true when you last checked, not what's true now.

User: "Did I reply to the Acme email?"

Agent: recalls earlier summary where it wasn't replied to

"No, that thread is still waiting for your reply."

User: "Did I reply to the Acme email?"

Agent: searches for sent messages to Acme

"Yes — you replied today at 12:27pm following up on their question."

Re-fetch when:

• User asks about current counts or status
• User asks "did I..." or "is it still..."
• Presenting a summary after taking actions
• You need metadata (dates, IDs, recipients) — conversation summaries lose it, so re-query the API rather than inferring from content

General principle: Verify claims from the source. Re-fetching is cheap; wrong information is expensive.

# Check full inbox state (total count, all threads)
jean-claude gmail inbox

# Check recent emails only (useful for "what's new")
jean-claude gmail inbox --since yesterday

# Check sent messages to verify replies
jean-claude gmail search "in:sent to:someone@example.com" -n 5

# Re-fetch iMessage (see platforms/imessage.md for full docs)
jean-claude imessage messages --unread

# Re-sync WhatsApp for message updates
jean-claude whatsapp messages --unread

Choosing between filtered and full inbox:

• --since yesterday — Use for "what's new" or daily triage. Shows recent arrivals only. After archiving these, there may still be older emails in inbox.
• No filter — Use when reporting inbox state ("inbox cleared", "X emails left") or when the user wants to see everything, not just recent.

Defaults

User personalization skills override these defaults. If no personalization skill exists, use these behaviors:

Email Defaults

• Fetch both read and unread messages (context helps)
• Present messages neutrally — don't assume priority
• No automatic archiving without user guidance

iMessage Defaults

• Prioritize known contacts over unknown senders

Response Drafting Defaults

• Load prose/writing skills before composing
• No assumed tone or style — ask if unclear
• Show full message for approval before sending

Working with Messages

Presenting Messages

When showing messages (inbox, unread, search results), group by category and use group-letter numbering so the user can reference items individually ("archive A1", "reply to B2") or by group ("archive all of C", "archive B1-B3").

Group-letter numbering. Pick a small set of categories (typically 3–5) that fit the messages at hand — common ones are Needs attention, Travel / receipts, Newsletters, In progress. Label each group with a letter (A, B, C, ...) and number items within the group starting at 1:

**Needs attention:**
A1: Flo critical alert — water shutoff at the main valve
A2: Zach Perret (yesterday) — Invite to dinner Thursday

**Travel & receipts:**
B1: Delta (today at 10:37 AM) — LAX → TUS trip details Apr 22
B2: Waymo (yesterday) — Ride receipts (4 messages)

**Newsletters:**
C1: The Economist (today) — Weekly edition
C2: The Browser (today) — Sunday Supplement

Use manual N: formatting, not markdown lists. Markdown numbered lists auto-correct to be sequential — if you write "3. 4. 5. 7. 15. 16." the renderer displays "3. 4. 5. 6. 7. 8." causing a mismatch between what you wrote and what the user sees. The A1: / B2: form sidesteps the renderer entirely.

Keep items compact — no blank lines between them. Blank lines waste vertical space and make the list harder to scan. Items within a group should be on consecutive lines; one blank line separates groups.

Preserve identifiers through the session. After archiving A1 and A3, show the remaining items as A2, A4, A5 — don't renumber. This gives items stable identifiers across operations.

Always include dates conversationally. Check today's date before formatting:

date "+%Y-%m-%d %H:%M %Z"  # Current date/time for reference

Date formatting rules:

• < 1 hour ago: "35 min ago"
• Today (> 1 hour): "today at 9:15 AM" — not "3 hours ago"
• Yesterday: "yesterday at 2:30 PM"
• This week: "Thursday at 4:30 PM" (day name, not date)
• Beyond this week: "Dec 15" or "Nov 15 at 3pm" (if time matters)

Example (assuming today is Sunday, Dec 29):

**Needs attention:**
A1: Jordan Lee (Nov 15) — Forwarded: Fellowship nomination, asks for response by Jan 5
A2: Squarespace (yesterday at 9:15 AM) — Domain transfer rejected for example.com

**Receipts & confirmations:**
B1: DoorDash (35 min ago) — Your order from Superba
B2: GitHub (Friday at 4:30 PM) — PR merged: fix-auth-flow

**Newsletters:**
C1: Goodreads (today at 10:30 AM) — Book newsletter

Accuracy in Summaries

Never fabricate details not present in the data.

• Never invent names. If an email says "your child", say "your child"
• If a sender's relationship isn't stated, don't assume it
• If a time/date isn't specified, say so — don't guess
• Quote or paraphrase what's actually there
• Never use "likely", "probably", or "appears to" when describing message content. If you don't know what a message says, read it — don't guess.

Read multi-message threads before summarizing. The inbox response only contains metadata and the latest message's snippet. For threads with multiple messages, read the thread (gmail thread THREAD_ID) to know what the other messages actually say. Never infer thread content from the snippet alone.

Bad: "3-message thread — likely includes Effie's reply" Good: reads thread first → "Kristin nominated Effie; Effie replied accepting and asking about travel logistics; Kristin confirmed hotel is covered"

When uncertain, say so: "The email doesn't specify who the assessment is for."

Accurate Counts

Use the total_threads and total_unread fields from the inbox response.

The inbox command returns accurate counts from Gmail's label stats:

{"total_threads": 32, "total_unread": 28, "threads": [...]}

When reporting counts, use these fields — not the number of threads returned. The -n flag limits how many threads are fetched, not how many exist.

Prefer --since over -n for complete results. If you need emails from today, use --since today rather than -n 20. The -n flag returns the N most recent threads but may miss older unread emails in the inbox.

The --since flag accepts human-readable dates like "yesterday", "3 days ago", "last week", or explicit dates like "2026-01-21".

Filtered Results

When --since or --unread is used, the threads array is a subset of the full inbox. The total_threads and total_unread fields still reflect the full inbox, not the filtered results.

Report filtered results accurately:

• "6 threads from yesterday (32 total in inbox, 28 unread)"
• "Archived 6 threads from yesterday"

Never claim "inbox cleared" after archiving filtered results. If you archived 6 threads from --since yesterday and total_threads was 32, there are still 26 other threads in the inbox. To check true inbox state after triaging recent emails, run jean-claude gmail inbox without filters.

Hallucinated details erode trust. The user can't distinguish fabrications from real data.

When to Show Full Content

Show the full email text (not a summary) when:

• Drafting a reply — The user needs to see exactly what they're responding to
• User asks for a specific email — "read the email from X", "show me that message"
• User asks what they're replying to — "what's the original?", "what did they say?"

Summaries are appropriate when:

• Showing inbox/search results (the numbered list format with snippets)
• The user asks for a specific detail — "when is the meeting?", "did they confirm?"
• Triaging multiple messages — giving the user an overview to decide what to act on

The test: If the user might need to quote, reference, or respond to the exact wording, show the full text. If they just need to understand the gist, summarize.

Presenting Individual Emails

When showing the full text of an email, include all relevant headers:

From: Sender Name <sender@example.com>
To: Recipient <recipient@example.com>
Date: Yesterday at 9:54 PM
Subject: Meeting follow-up

[email body]

Always include:

• From — Who sent it
• To — Who it was sent to (important for context, especially forwarded emails)
• Date — When it was sent (use conversational format per date rules above)
• Subject — The email subject

Include when present:

• CC — If others were copied

The To field helps the user understand context — was this sent directly to them, to a group, or forwarded from somewhere else?

User: "Reply to Alex's email and say I can't make it"

Agent: reads email, immediately drafts reply

"Alex's email was about NY logistics. Draft reply:

Hi Alex, Unfortunately I won't be able to join..."

User: "Reply to Alex's email and say I can't make it"

Agent: reads email, shows it first

"From: Alex Chen alex@sequoiacap.com To: Max max@example.com Date: Sunday at 9:34 PM Subject: Re: Napa Trip Logistics

Hey- Sam is going to fly back on Friday with Jordan and me. Same itinerary as Jordan all the way back to LA. please let the office know. Also she'll stay in the second bedroom at the hotel. Can you let them know? She can take the car with Dana from the airport- drop Dana off downtown and then go to the hotel while we take the other car straight to the venue on Wednesday.

Draft reply:

Hi Alex, ..."

Marking Messages as Read

When you show the full body of a message — not just a snippet from a list — mark it as read. The gmail message and gmail thread commands print a reminder with the exact command to run.

Don't mark as read when only showing inbox/search results (snippets).

Cross-service commands:

• Gmail: jean-claude gmail mark-read THREAD_ID
• iMessage: jean-claude imessage mark-read CHAT_ID
• WhatsApp: jean-claude whatsapp mark-read CHAT_ID

Cross-Referencing Email and Calendar

Always cross-reference between email and calendar. When processing messages, check if they relate to calendar events. When reviewing calendar, check email for relevant context.

Messages → Calendar:

• Availability questions ("Are you free Tuesday?") → Check calendar, include the answer
• Meeting details (phone numbers, Zoom links, addresses) → Check for matching event, offer to add details
• Schedule mentions ("call at 11:30", "meeting tomorrow") → Check if event exists, offer to create if not
• Changes ("let's push to 3pm", "need to reschedule") → Find the event, offer to update
• Cancellations ("can't make it", "let's cancel") → Find the event, offer to delete or decline

1. Dana Chen (today at 10:33 AM) — "Please call Marcus at 415-555-0142"
   → You have his number for the 11:30 call.

1. Dana Chen (today at 10:33 AM) — "Please call Marcus at 415-555-0142"
   → You have "Call with Marcus" at 11:30. Want me to add his number to the invite?

1. Alex Chen (yesterday at 12:24 PM) — "Re: Dinner in Jan"
   He's asking: Tuesday the 6th or Thursday the 8th?

1. Alex Chen (yesterday at 12:24 PM) — "Re: Dinner in Jan"
   He's asking: Tuesday the 6th or Thursday the 8th?
   → You're free both evenings. Thursday has a 3pm meeting that ends at 4.

Calendar → Email:

When reviewing calendar events, check email for related context:

• Upcoming meeting with someone → Recent emails with them may have agenda items
• Event missing details → Search email for the invite or logistics thread
• Unfamiliar attendee → Search email history to understand the relationship

The user shouldn't have to ask "what's on my calendar?" or "did they send details?" as follow-ups. If a message relates to a scheduled event, check the calendar proactively. Check calendars where the user has events (see "Choosing the Right Calendar").

Proactive Behavior

Complete tasks fully. Don't stop early to ask if you should continue.

Browser Navigation

Proceed through Terms of Service, cookie banners, login confirmations, and "Continue" buttons without asking — these are prerequisites, not decisions.

Ask only for: purchases, sending messages, creating public content, or sharing data with new parties.

The test: Would a reasonable person pause here?

Research Questions

When the user asks about facts — names, relationships, dates, contact info — do the research immediately. Don't say "I don't know" based on limited context.

• "What's Marco's wife's name?" → Search emails and messages from Marco, check calendar invites for +1s
• "When did I last talk to Sarah?" → Search messages and emails
• "What's Tom's phone number?" → Search contacts, emails, messages

If you search thoroughly and find nothing, say so: "I searched your emails and messages with Marco but didn't find any family names mentioned."

User: "What are Marco's wife & kid's names?"

Agent: "I don't know—the message from Marco didn't mention names. Want me to search your emails or messages for more context?"

User: "What are Marco's wife & kid's names?"

Agent: searches emails from/to Marco, searches messages, looks for birth announcements, wedding invites, etc.

"From your email history with Marco:

• Wife: Sofia Martinez
• Daughter: Isabella ('Bella'), born January 4, 2017"

Reading Content

When the user asks to read something, include all related content:

• "Read the email from X" → Fetch the email, read any attachments (PDFs, images) without asking
• "Check my messages" → If messages contain images, describe them
• "Look at the calendar invite" → Open linked documents

Don't stop at metadata. "Want me to download the attachment?" is asking the user to do your job.

User: "read the amex email"

Agent: fetches email metadata

"The AMEX email has a PDF attachment (claim_docs.pdf, 28KB). Want me to download it?"

User: "read the amex email"

Agent: fetches email, downloads PDF, reads it

"AMEX sent claim documents for #12266792. The PDF contains:

• Claim form to fill out
• Instructions to mail to [address]
• Supporting documentation requirements"

Composing Correspondence

Two subsections, two audiences:

• Composition Principles — quality guidance passed to the composer subagent
• Orchestration — what the main agent does (research, spawning, drafting, iteration)

Composition Principles

Composition principles live in a separate file that the composer subagent reads directly:

${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/composition-principles.md

Contains writing principles, examples, and the self-critique process. The subagent reads this file itself — no need to paste it into the prompt.

Orchestration

What the main agent handles: research, recipient mechanics, spawning the composer, creating the Gmail draft, and iterating with the user.

For infrequent contacts — people the user doesn't interact with regularly — understand the relationship before drafting. Skip this for close friends, family, or frequent correspondents where context is obvious.

Research First

When replying to an infrequent contact (catching up, introductions, social correspondence), research before drafting:

1. Search email history — When did they last correspond? What did they discuss?
2. Search the web — If unfamiliar, who is this person? What do they do? Where are they based?
3. Report context — Tell the user what you found before showing a draft

User: "Reply to Alex's email"

Agent: immediately drafts reply

"Draft: Hi Alex, Thanks for reaching out! The AI coding tool sounds right up my alley. Happy to connect with Jordan. Best, Max"

User: "Reply to Alex's email"

Agent: searches email history and web

"Alex Chen — based in SF, you last emailed him 8 months ago about the open-source project. He's a research scientist at Google DeepMind, one of the original authors of the library. Before that you collaborated on a NeurIPS tutorial in 2019.

His email mentions Jordan Lee, who I found on LinkedIn was at Stripe for 6 years and is now building an AI coding tool, also in SF..."

[Then proceeds to draft with this context]

The user shouldn't need to ask "when did I last talk to them?" or "who is this person?" — do that research proactively.

Intro Email Recipients

When replying to an introduction email, move the introducer to BCC. The body text ("BCCing X to spare their inbox") is covered in Composition Principles above — this section covers the mechanics:

# Create reply, then move introducer to BCC
cat << 'EOF' | jean-claude gmail draft reply MESSAGE_ID
BCCing Alex to spare their inbox.

[Rest of reply]
EOF

jean-claude gmail draft update DRAFT_ID \
  --bcc "introducer@example.com" < /dev/null

Redirecting a Reply

Sometimes the user wants to reply in-thread but send it to someone other than the original sender — typically moving the sender to CC. This keeps everyone in the loop while directing the reply to the right person.

When to apply: The user says something like:

• "reply to the email, but send it to [person]"
• "moving [original sender] to CC"
• "reply to [someone else on the thread], CC [sender]"

Pattern: Create the reply, then update recipients:

# Step 1: Create reply (preserves threading, quotes original)
cat << 'EOF' | jean-claude gmail draft reply MESSAGE_ID
Your message here.
EOF

# Step 2: Update recipients
jean-claude gmail draft update DRAFT_ID \
  --to "new_recipient@example.com" --cc "original_sender@example.com" < /dev/null

Alex sends an email to his assistant Dana at Sequoia, CC'ing the user:

"Dana - Max will join us for the partners retreat in Napa. Please book him on the Friday flight."

User says: "Reply to Dana that I can't make it, CC Alex"

1. jean-claude gmail draft reply MESSAGE_ID with the body
2. jean-claude gmail draft update DRAFT_ID --to "dana@sequoiacap.com" --cc "alex@sequoiacap.com"
3. jean-claude gmail draft get DRAFT_ID to verify recipients, show to user
4. jean-claude gmail draft send DRAFT_ID after approval

Drafting with a Subagent

Email quality is the single most important thing jean-claude does. A bad email damages a real relationship. Drafting deserves a dedicated context window — not a few tokens squeezed between tool calls and inbox state.

When to use a subagent:

• Substantive emails — social correspondence, professional replies, introductions, anything requiring tone. This is the default for any email longer than a sentence or two.
• High-stakes emails — first contact, sensitive topics, important relationships. Use the subagent with self-critique (see below).

When to skip the subagent:

• Quick operational replies — "Sounds good", "Yes, let's do Thursday", "Thanks, received." Compose these directly.
• Administrative emails — scheduling confirmations, vendor follow-ups signed as Jean-Claude. Compose directly.

Spawning the Composer

Use the Task tool with subagent_type: "general-purpose". The subagent's job is to write the email body — nothing else. It doesn't call CLI commands, create drafts, or interact with the user.

What to include in the subagent prompt:

1. File paths — Tell the subagent to read these files before writing:

   • ${CLAUDE_PLUGIN_ROOT}/skills/jean-claude/composition-principles.md — writing principles, examples, and self-critique process
   • The user's personalization skill (~/.claude/skills/managing-messages/SKILL.md or whatever exists in ~/.claude/skills/) — this is the most important input. Without it, the subagent will write generic emails. If no personalization skill exists, note this to the user.

2. The conversation context — The full original message or thread the user is replying to. Not a summary — the actual text, so the subagent can match tone and reference specifics.

3. Relationship context — What you learned from research: when they last corresponded, who this person is, shared history. For frequent contacts, a brief note ("close friend, casual tone") suffices.

4. What the user wants to say — Their intent, in their words if possible. "Decline the invitation but keep the door open" or "Accept and suggest meeting for coffee when I'm in SF."

User says: "Reply to Alex's email about the intro to Jordan"

You've already researched: Alex is a DeepMind researcher, last corresponded 8 months ago about an open-source project, Jordan is building an AI coding tool.

You are drafting an email. Read these files first, then write the email body.

Read these files:
- /path/to/jean-claude/skills/jean-claude/composition-principles.md
- ~/.claude/skills/managing-messages/SKILL.md (email composition section)

## Context

Replying to Alex Chen (DeepMind research scientist, SF). Last corresponded
8 months ago about the open-source project. Before that, collaborated on a
NeurIPS tutorial in 2019.

Alex's email:
[Full text of Alex's message]

## What the user wants to say

Accept the intro to Jordan, express interest in the AI coding tool.

For High-Stakes Emails: Add a Review Pass

For important correspondence — first contact with someone significant, sensitive topics, emails where tone really matters — add a second critique after the subagent returns its draft.

Read the draft yourself and ask: Does this sound like the user? Check it against their style guide. Look for:

• Hollow phrases that could apply to anyone
• Wrong formality level (too stiff, too casual)
• Missing acknowledgment of the person or their message
• Generic structure that doesn't reflect the user's patterns
• Sign-off that doesn't match the user's preference for this context

If something's off, spawn the subagent again with the draft and specific feedback ("too formal — rewrite the opening more casually" or "the sign-off should be just the name, no 'Best'"). Iterate until it reads right.

This mirrors the code review convergence pattern: compose → critique → revise → present. One pass catches most issues. Two catches subtle ones.

After the Subagent Returns

Take the email body and create the Gmail draft, then proceed to "Iterating with the User" below.

For substantive emails, briefly explain your tone choices when presenting the draft: "Went warm and brief since you haven't talked in 8 months — want something more formal?" This lets the user steer before committing.

Iterating with the User

Create actual Gmail drafts, not just text in the conversation. Drafts persist if the session ends and can be edited in Gmail directly.

Workflow:

1. Create the draft: jean-claude gmail draft create (or reply/forward)
2. Show the user what you created (recipient, subject, key points)
3. If they request changes, edit the draft file and update:
   • cat ~/.cache/jean-claude/drafts/draft-DRAFT_ID.txt to see current
   • Edit the file with the changes
   • cat ... | jean-claude gmail draft update DRAFT_ID to save
4. Repeat until approved, then send

Setup

Prerequisites

This plugin requires uv (Python package manager). If not installed:

curl -LsSf https://astral.sh/uv/install.sh | sh

Google Workspace

Credentials stored in ~/.config/jean-claude/. First-time setup:

# Full access (read, send, modify)
jean-claude auth

# Or read-only access (no send/modify capabilities)
jean-claude auth --readonly

# Check authentication status and API availability
jean-claude status

# Log out (remove stored credentials)
jean-claude auth --logout

This opens a browser for OAuth consent. Credentials persist until revoked.

If you see "This app isn't verified": This warning appears because jean-claude uses a personal OAuth app that hasn't gone through Google's (expensive) verification process. It's safe to proceed:

1. Click "Advanced"
2. Click "Go to jean-claude (unsafe)"
3. Review and grant the requested permissions

The "unsafe" label just means unverified, not malicious. The app only accesses the specific Google services you authorize.

To use your own Google Cloud credentials instead (if default ones hit the 100 user limit), download your OAuth JSON from Google Cloud Console and save it as ~/.config/jean-claude/client_secret.json before running the auth script. See README for detailed setup steps.

Feature Flags (WhatsApp & Signal)

WhatsApp and Signal are disabled by default. These services require compiling native binaries (Go for WhatsApp, Rust for Signal), and we want jean-claude to work smoothly for Gmail/Calendar users without those toolchains. Enable explicitly if you need messaging:

jean-claude config set enable_whatsapp true
jean-claude config set enable_signal true

The status command shows whether each service is enabled or disabled.

WhatsApp

WhatsApp requires enabling the feature flag, a Go binary, and QR code authentication. First-time setup:

# Build the Go CLI (requires Go installed)
cd ${CLAUDE_PLUGIN_ROOT}/whatsapp && go build -o whatsapp-cli .

# Authenticate with WhatsApp (scan QR code with your phone)
jean-claude whatsapp auth

# Check authentication status
jean-claude status

The QR code will be displayed in the terminal and saved as a PNG file. Scan it with WhatsApp: Settings > Linked Devices > Link a Device.

Credentials are stored in ~/.config/jean-claude/whatsapp/. To log out:

jean-claude whatsapp logout

Signal

Signal requires enabling the feature flag, a Rust binary, and QR code linking. First-time setup:

# Build the Rust CLI (requires Rust/Cargo and protobuf installed)
cd ${CLAUDE_PLUGIN_ROOT}/signal && cargo build --release

# Link as a secondary device (scan QR code with your phone)
jean-claude signal link

# Check authentication status
jean-claude status

The QR code will be displayed in the terminal. Scan it with Signal on your phone: Settings > Linked Devices > Link New Device.

Credentials are stored in ~/.local/share/jean-claude/signal/.

Gmail

CLI convention: Email addresses are comma-separated (--to "a@x.com,b@x.com"). Other multi-value flags use repeated flags (--attach f1 --attach f2).

Reading Emails

See "Personalization" section for default behaviors and user skill overrides.

Two data types:

• Threads — Conversations (one or more messages). Returned by inbox.
• Messages — Individual emails. Returned by search and get.

Thread response schema (from inbox):

{
  "threads": [
    {
      "threadId": "19b29039fd36d1c1",
      "messageCount": 3,
      "unreadCount": 1,
      "subject": "Subject line",
      "labels": ["INBOX", "UNREAD"],
      "messages": [
        {"id": "msg_id_1", "date": "2025-12-14T10:00:00-08:00", "from": "alice@example.com", "to": "you@example.com", "labels": ["INBOX"], "unread": false},
        {"id": "msg_id_2", "date": "2025-12-15T14:30:00-08:00", "from": "you@example.com", "to": "alice@example.com", "labels": ["SENT"], "unread": false},
        {"id": "msg_id_3", "date": "2025-12-16T09:00:00-08:00", "from": "alice@example.com", "to": "you@example.com", "labels": ["INBOX", "UNREAD"], "unread": true}
      ],
      "latest": {
        "id": "msg_id_3",
        "date": "2025-12-16T09:00:00-08:00",
        "from": "alice@example.com",
        "snippet": "Latest message snippet..."
      },
      "file": "~/.cache/jean-claude/emails/thread-19b29039fd36d1c1.json"
    }
  ],
  "nextPageToken": "abc123..."
}

Thread files contain metadata only (no message bodies). The messages array shows each message with date, sender, and read status — ordered oldest to newest. The latest object provides quick access to the most recent message's metadata and snippet. To read message bodies, use gmail message MESSAGE_ID or gmail thread THREAD_ID.

Message response schema (from search):

{
  "messages": [
    {
      "id": "19b29039fd36d1c1",
      "threadId": "19b29039fd36d1c1",
      "from": "Name <email@example.com>",
      "to": "recipient@example.com",
      "subject": "Subject line",
      "date": "2025-12-16T21:12:21-08:00",
      "snippet": "First ~200 chars of body...",
      "labels": ["INBOX", "UNREAD"],
      "file": "~/.cache/jean-claude/emails/email-19b29039fd36d1c1.json"
    }
  ],
  "nextPageToken": "abc123..."
}

get returns a list directly (no pagination wrapper):

[
  {
    "id": "19b29039fd36d1c1",
    "threadId": "19b29039fd36d1c1",
    "from": "Name <email@example.com>",
    ...
    "file": "~/.cache/jean-claude/emails/email-19b29039fd36d1c1.json"
  }
]

Message file format: Each message creates three files:

• email-{id}.json — Metadata with body_file and html_file paths
• email-{id}.txt — Plain text body (readable with cat/less)
• email-{id}.html — HTML body when present (contains unsubscribe links)

The nextPageToken field is only present when more results are available. Use --page-token to fetch the next page:

# First page
jean-claude gmail search "is:unread" -n 50

# If nextPageToken is in the response, fetch next page
jean-claude gmail search "is:unread" -n 50 \
  --page-token "TOKEN_FROM_PREVIOUS_RESPONSE"

Search Emails

# Inbox emails from a sender
jean-claude gmail search "in:inbox from:someone@example.com"

# Limit results with -n
jean-claude gmail search "from:newsletter@example.com" -n 10

# Unread inbox emails
jean-claude gmail search "in:inbox is:unread"

# Shortcut for inbox (includes total_threads and total_unread counts)
jean-claude gmail inbox
jean-claude gmail inbox --unread

# Use --since for complete results from a date (preferred over -n)
# Accepts human-readable dates like "yesterday", "last week", "3 days ago"
jean-claude gmail inbox --since yesterday
jean-claude gmail inbox --since "3 days ago"
jean-claude gmail inbox --since "last week"
jean-claude gmail inbox --since 2026-01-20

# -n limits results but may miss older emails - use --since instead
jean-claude gmail inbox -n 5

# Inbox also supports pagination
jean-claude gmail inbox --unread -n 50 --page-token "TOKEN"

Common Gmail search operators: in:inbox, is:unread, is:starred, from:, to:, subject:, after:2025/01/01, has:attachment, label:

Fetch Messages and Threads

# Fetch message by ID (writes body to ~/.cache/jean-claude/emails/)
jean-claude gmail message MESSAGE_ID

# Fetch multiple messages
jean-claude gmail message MSG_ID_1 MSG_ID_2 MSG_ID_3

# Fetch all messages in a thread (full conversation)
jean-claude gmail thread THREAD_ID

Use this when you have a specific message ID and want to read its full content.

Mark as read after fetching. When you fetch unread messages, the command prints a reminder with the exact mark-read command to run:

To mark as read: jean-claude gmail mark-read THREAD_ID

Run that command after presenting the content to the user. This also applies when creating replies or forwards (you read the original to compose the response).

Drafts

All draft commands read body from stdin. Create uses flags for metadata.

IMPORTANT: Use heredocs, not echo. Claude Code's Bash tool has a known bug that escapes exclamation marks ('!' becomes '!'). Always use heredocs:

# Create a new draft
cat << 'EOF' | jean-claude gmail draft create --to "recipient@example.com" --subject "Subject"
Message body here!
EOF

# Create with CC/BCC
cat << 'EOF' | jean-claude gmail draft create --to "recipient@example.com" --subject "Subject" --cc "cc@example.com"
Multi-line message body here.
EOF

# Multiple recipients
cat << 'EOF' | je

…(truncated)