/announcement-digest

name: announcement-digest
description: "Compile a digest of announcement emails from a configured source. Use for recurring school, employer or organization email summaries."

Announcement Digest

Adapt for your source — swap the @[SOURCE] label, names, and classification rules.

This is a sanitized version of a working school-email digest skill.

Full walkthrough: https://claudeblattman.com/workflows/school-digest/

*v1.0 — Announcement email digest skill*

Compile a digest of high-volume announcement emails from a single source (school, employer, nonprofit, professional association). Extract what matters, skip the noise, send a summary.

Prerequisites

• **Gmail MCP** — for searching and reading emails
• **Gmail label + filter** — all source emails auto-labeled (e.g., `@School`, `@CompanyNews`)
• **Classification rules** — a config file with your calibrated rules (see walkthrough)
• **send-email.py** or equivalent — for sending the compiled digest

Config & State Files

• **Rules:** `[YOUR_CONFIG_PATH]/digest-rules.md`
• **State:** `[YOUR_STATE_PATH]/digest-state.json`

State file tracks: `last_digest_date`, `digest_count`, `reported_message_ids`, `reported_message_ids_ttl_days` (default 30).

Execution Steps

Step 1: Read State & Check Frequency

1. Read `digest-state.json`. 2. If `last_digest_date` is not null and fewer than 2 days have passed:

• Output: "Nothing new since [last_digest_date]. Next digest available [date]."
• **Exit.**

3. Store `digest_count` (needed for first-time intro check at Step 5).

Step 2: Fetch Labeled Emails

1. Compute lookback date: `last_digest_date` if set, otherwise 14 days ago. 2. Search Gmail: `label:[YOUR_LABEL] after:[lookback_date in YYYY/MM/DD]`

• **Do NOT use `is:unread`** — labeled emails may or may not be marked read. Query by date range.
• Page through all results (follow `next_page_token`).

3. Filter out any message IDs already in `reported_message_ids`. 4. If no new emails remain, exit with message.

Step 3: Fetch Email Content

1. For each new message, fetch full content via `get_gmail_message_content`.

• Fetch in parallel batches of 10.

2. Group messages by Gmail thread ID. Keep only the most recent non-you message per thread. 3. Accumulate total word count across all fetched message bodies. 4. **Prune `reported_message_ids`**: Remove IDs older than `reported_message_ids_ttl_days`.

Step 4: Classify Each Email

Read your rules config for the full rule set. For each email (or thread), apply classification in this order:

4a. Skip Check

Apply SKIP rules first. If any match, mark as `skip` and record `skip_reason`.

4b. Source Assignment

Determine which sub-category the email belongs to (e.g., which child's school, which department, which program).

4c. Urgency Classification

Apply rules from config using this priority order: 1. **Sender pattern override:** Personal sender (not `noreply@`) → HIGH 2. **HIGH keyword match** in subject + first 200 chars → HIGH 3. **LLM extraction** for everything else:

> Extract from this email. Return as JSON. > (1) category: action_required / calendar_event / fyi / skip > (2) urgency: high / medium / low > (3) summary: 1-2 sentences of what you need to know > (4) deadline: specific date/time if any, else null > (5) deadline_quote: verbatim text from email stating the date/deadline — copy exact wording, do not paraphrase. null if no deadline. > (6) skip_reason: if skip, one phrase why

4. **HIGH keyword overrides LLM urgency** upward (never downward). 5. **SKIP keyword overrides to skip** unless LLM said `action_required`.

Step 4.5: Write Classified Artifact

Create a private temp directory with `mktemp -d` and write full classified data to `digest-classified.json` inside that directory. This serves as:

• A **compaction safeguard** — if context is lost, re-compose from this file
• A **debug artifact** — review classification accuracy after the fact

Step 5: Compose Digest

Format as plain text. Structure:

[YOUR SOURCE] EMAIL DIGEST
[date range]
Compressed ~[W] words across [N] emails ([M] threads) into this digest.
──────────────────────────────

[CATEGORY 1]

▶ ACTION ITEMS
  • [summary] — deadline: [date]
    Source: "[deadline_quote]"

📅 UPCOMING
  • [event/date]

ℹ️ FYI
  • [1-line summary]


[CATEGORY 2]

▶ ACTION ITEMS
  ...

+ [N] lower-priority items not shown
──────────────────────────────

**Rules:**

• Action items (HIGH) first, deadlines noted inline
• For HIGH items with a `deadline_quote`, include: `Source: "[deadline_quote]"` — makes factual errors visible
• Calendar events (MEDIUM) second
• FYI (LOW) third, 1 line each
• Omit empty sections
• **First-time send** (if `digest_count` was 0): Prepend intro explaining what this automated digest is and who it's from.

Step 6: Save State & Send

1. **Write state file FIRST** (before sending):

• Add ALL processed message IDs to `reported_message_ids` (including skipped)
• Set `last_digest_date` to today
• Increment `digest_count`

2. **Draft the digest** via your preferred method. Ask for explicit approval before sending. The public template must not auto-send.

3. **Do NOT mark source emails as read.** The digest is additive — it doesn't replace checking the source.

Step 7: Log Performance

Log tool call count, duration, and volume for future optimization.

Edge Cases

• **No emails:** Exit with message at Step 2.
• **All already reported:** Exit with message at Step 2.
• **Send failure after state write:** State is already saved (prevents duplicate digests). Log the error and retry manually.

What This Skill Does NOT Do

• Does not mark emails as read
• Does not apply Gmail labels
• Does not modify any emails

Building Your Own Rules Config

The classification rules are the hardest part. Don't try to write them upfront — calibrate from real data:

1. Pull 20 recent emails from your source 2. Rate each: action_required / calendar / fyi /