name: brand-monitoring-automation description: Build and operate brand monitoring pipelines that fetch news/social sources, score/filter mentions, generate reports, and send only final digests to Telegram.

Brand Monitoring Automation

Use this skill when maintaining a brand monitoring system that combines RSS/news collection, social keyword scraping, relevance filtering, weekly reporting, and Telegram digest delivery.

Core workflow

1. Fetch sources into a local DB silently.
2. Normalize, dedupe, score relevance, and classify sentiment.
3. Apply a strict time window before processing.
4. Generate one final digest for Telegram and one report file for weekly review.
5. Keep intermediate steps off the user-facing thread.

Important rules

• Only the final announcement should be posted to the Telegram monitoring thread.
• Fetch, filter, and scoring steps must run in the background with no thread updates unless explicitly requested.
• DANGER: Do not rely on LLM memory for dates. BẮT BUỘC use execute_code to calculate dynamic search windows (e.g., after:YYYY-MM-DD) to avoid pulling ancient (2016-2021) or "fabricated" news.
• Format: Every digest item MUST include [Ngày đăng], a concise impact summary, a clickable URL source, and a clear Tier/Risk classification.
• Weekly reports should be written to disk first, then announced via Tailscale Funnel link.
• A delivery test is not a placeholder ping. A real test must run the same end-to-end path as the scheduled job: fetch the relevant window, filter and dedupe, build the final digest, then send only that final digest.
• If a user moves the workflow to an official automation blueprint, remove or disable any older hand-built Kanban boards, cron jobs, and scripts that would duplicate the same output.
• Do not assume a named automation blueprint exists just because a prompt or conversation names it. Verify the slash command, blueprint definition, or cron job record before claiming it is installed.

Time-window guidance

For cadence-based jobs, use a window that matches the schedule (e.g., 72 hours for a 3-day cycle). Use execute_code to verify the delta.

Telegram delivery guidance

• Send a single final digest rather than per-item chatter.
• Use Tailscale Funnel URLs (https://vmi3427693.tail8c1aaf.ts.net/files/...) for all reports.
• Never route cron stdout/stderr to the public monitoring thread (set deliver: local in cron).

Social scraping guidance

• Keyword-based social collection should be scoped to the business context, not broad brand name matching alone.
• Prefer context filters such as product, program, location, or institutional names when brand terms are overloaded.
• If public search surfaces are noisy, narrow by geography and institutional context before scoring.

Pitfalls

• A fixed 72-hour window can be valid for catch-up, but it is not automatically correct for every cadence.
• Broad brand keywords often pull in global or unrelated communities.
• Cron jobs with origin delivery can leak intermediate logs into the wrong thread.
• Do not post per-source or per-item updates to the Telegram thread when the user asked for only the final announcement.

Verification

Before marking the workflow ready:

• Confirm the digest target is the correct Telegram chat/thread.
• Confirm the cron job delivery mode does not emit intermediate logs to the thread.
• Confirm the report file is written successfully before announcement.
• Confirm the live digest only contains the intended time window and context-specific items.

Support files

• references/telegram-thread-routing.md — thread delivery lessons and cron delivery pitfalls.
• scripts/test_digest_window.py — deterministic check for date-window boundaries and digest scoping.