slack-respond

Slack inbox triage and response drafting tool. Pulls recent messages needing Winston's attention, summarizes them for quick scanning, then drafts replies in his voice.

Arguments

• $ARGUMENTS (optional):
  • (empty) — last 24 hours (default)
  • 15m, 30m — relative time window (any number + m for minutes)
  • 2h, 4h, 8h — relative time window (any number + h for hours)
  • today — since midnight
  • yesterday — yesterday's messages

Step 1: Prerequisites

1. Verify Slack MCP is connected by running a lightweight test call:

   mcp__slack__slack_search_channels with query: "general", limit: 1
   

   If this fails, tell the user:

   Slack MCP is not connected. Add it with: claude mcp add slack --transport http --scope user <slack-mcp-url> Then stop.

2. Read the Slack writing style guide into context:

   Read file: /Users/winstonhuang/Library/Mobile Documents/com~apple~CloudDocs/Documents/Claude Code/Slack/CLAUDE.md
   

   If the file is missing, warn the user but continue with these defaults: warm, direct, lowercase in DMs, !! for enthusiasm, short fragments, no formal sign-offs.

Step 2: Compute Time Window

Parse $ARGUMENTS into a Unix timestamp using Bash date commands (macOS/BSD syntax). Use the after parameter (Unix timestamp) on the search tool instead of after:YYYY-MM-DD in the query string — this gives true sub-day precision.

Argument	Bash command
Nm (e.g. 15m, 30m)	date -v-Nm +%s
Nh (e.g. 2h, 4h, 8h)	date -v-Nh +%s
(empty) or 24h	date -v-24H +%s
today	date -j -f "%Y-%m-%d" "$(date +%Y-%m-%d)" +%s (midnight today)
yesterday	date -j -f "%Y-%m-%d" "$(date -v-1d +%Y-%m-%d)" +%s (midnight yesterday)

Store the computed timestamp as $UNIX_TS for use in the after parameter on search queries.

Also compute a human-readable label for the time window (e.g. "last 24 hours", "last 15 minutes", "last 2 hours", "today", "yesterday") for use in the summary header.

Step 3: Run 3 Parallel Searches

Use mcp__slack__slack_search_public_and_private for all three searches. Run them in parallel since they are independent. Use sort: "timestamp" and sort_dir: "desc" for all.

1. Mentions: query: "<@U09A00TLNV6>", after: "$UNIX_TS" — catches @-mentions across all channels
2. DMs: query: "to:me", after: "$UNIX_TS" — direct messages to Winston
3. Group DMs: query: "to:me", after: "$UNIX_TS" — group DM messages to Winston

If any search returns 20 results (the limit), paginate by calling again with page: 2, etc., until fewer than 20 results are returned.

De-duplicate results across all three searches by (channel_id, message_ts) pair — the same message may appear in multiple searches.

Step 4: Filter & Classify

Filter out:

• Messages FROM Winston (user ID U09A00TLNV6) — we only want messages TO or mentioning him
• Bot messages that are purely informational (e.g. deploy notifications, automated alerts) unless they contain a question or action request

Classify each remaining message as one of:

Needs Response

A message needs a response if ANY of these are true:

• Contains a question mark (?)
• Contains request verbs: "can you", "could you", "would you", "please", "need you to", "help with"
• Is a DM where Winston has NOT replied yet in the thread/conversation
• Is a direct @-mention with action language ("review", "check", "approve", "update", "look at", "thoughts?")

FYI Only

• Announcements, bot messages, informational shares
• Messages Winston has already replied to
• Channel-wide broadcasts not specifically requesting action from Winston
• Reactions-only or emoji-only messages

Step 5: Present Summary

Before presenting the summary, use mcp__slack__slack_read_thread (or slack_read_channel for DMs without threads) to pull full conversation context for every "Needs Response" item. The user needs enough context to decide what to say — a 1-line summary is not enough.

Present the inbox summary in this format:

## Slack Inbox — [time window label]

### Needs Response (N)

**1. #channel-name — @sender, 2h ago**
> [Full context: what the conversation is about, what was said before, what specifically the sender is asking/needing from Winston, and current status — e.g. whether others have already chimed in or if it's been resolved. 2-5 sentences.]

**2. DM from @person — 45m ago**
> [Full context as above. Include the sender's exact question or request when possible. Note the conversation history — what Winston last said, what they replied, what's still open.]

### FYI Only (N)
**3. #channel-name — @sender, 1h ago**
> [1-2 sentence summary — these don't need as much detail since no response is needed]

---
Which messages would you like to respond to? (numbers, "all", or skip)

Important: The goal is that Winston can read each summary and immediately dictate response intent ("for 1, say X") without needing to open Slack or ask for more context. Err on the side of too much detail rather than too little. Include:

• What the conversation is about (the topic/project)
• What the sender specifically needs from Winston
• What Winston has already said in the thread (if anything)
• Whether the issue appears resolved or still open
• Any relevant replies from others

If no messages are found at all, say:

Inbox zero! No messages needing attention in the [time window]. Try a broader window: /slack-respond 30m, /slack-respond 8h, or /slack-respond yesterday

Then stop.

Step 6: Draft Responses (Interactive)

When the user provides response intent (e.g. "for 1 say yes, for 2 tell them I'll check tomorrow"):

For each message the user wants to respond to:

1. Get full context: Use mcp__slack__slack_read_thread with the message's channel_id and thread_ts (or message_ts if not in a thread) to read the full conversation context.

2. Draft the response applying the Slack style guide from Step 1:

   • Match formality to context (DM = very casual lowercase, channel = casual-professional)
   • Use Winston's signature patterns: !!, --, fragments, "ofc", "yall", "prob", "lemme"
   • Keep DM responses to 1-2 short messages
   • Channel replies can be slightly longer but still concise
   • NEVER sound like an AI wrote it

3. Present the draft to the user:

   **Reply to #channel / @person:**
   > [drafted message]
   
   Send as draft? (y/n/edit)
   

4. On approval: Use mcp__slack__slack_send_message_draft with:

   • channel_id: the channel/DM where the message was
   • thread_ts: the thread timestamp (to reply in-thread)
   • message: the drafted text

5. Draft collision handling: If slack_send_message_draft fails because a draft already exists for that channel:

   • Warn the user: "A draft already exists in this channel. Options: (a) send directly with slack_send_message, (b) skip, (c) copy to clipboard"
   • If user chooses direct send, use mcp__slack__slack_send_message with explicit user approval
   • For multiple threads in the same channel: draft the first, then present remaining as text for the user to send manually or sequentially after the first draft is sent

6. Report the draft link for each successfully created draft.

Step 7: Wrap Up

After all responses are processed:

• Report total count of drafts created
• Note any skipped items or errors
• If there were FYI items, briefly note them as acknowledged

Error Handling

Scenario	Action
Slack MCP not connected	Print setup instructions and stop (Step 1)
No messages found	"Inbox zero!" + suggest broader time window
Draft already exists for channel	Warn + offer direct send or skip (Step 6.5)
Style guide not found	Warn + use general voice defaults (Step 1.2)
One-draft-per-channel limit	Draft first reply, present rest as text for manual send
Search pagination needed	Auto-paginate until fewer than 20 results returned
Thread read fails	Show original message snippet, draft based on that context alone