brand-mention-monitor

<skill name="brand-mention-monitor"> <overview> Watch what's being said about the brand across the open web without turning into either silence or an alert firehose. A daily cron re-prompts a persistent session that searches news, social platforms, and forums for {{brand_terms}}, dedupes each candidate against the ledger of mentions already reported, classifies sentiment and notability, and posts one digest to {{slack_channel}} — a one-line rollup for routine mentions, a full callout with a suggested response draft for anything negative or notable.

Proactive and schedule-driven; read-only against the public web, with a single Slack channel as the only output. It never replies, comments, or publishes anywhere — a suggested response is always a draft for a human to send. </overview>

<when-to-load> - The daily cron fires the brand-monitor run. - A human asks for a manual sweep of current brand mentions. - The brand terms or watch scope need updating (a new alias, product name, or common misspelling starts showing up). </when-to-load> <workflow>

Step 0 — Orient and resume

cat memory/brand-monitor-log.md 2>/dev/null || echo "(no ledger yet — first run)"

Read every mention already reported before searching. If this is the first run ever, there is nothing to dedupe against — search, classify, log everything, and report normally.

Step 1 — Load the brand terms

{{brand_terms}} is the source of truth for what counts as a mention this run: brand name, aliases, product names, and common misspellings, one per line. Search for exactly these terms — no more, no less.

Step 2 — Search the web

Use the sandbox's built-in web search across news sites, social platforms (X/Twitter, LinkedIn, Reddit, etc.), and forums (Reddit threads, Hacker News, industry forums). Run one search per term in {{brand_terms}}, plus a news-specific pass. No login, no connector credential — public search only.

Step 3 — Fetch and normalize each candidate

Fetch the full text (or a substantial excerpt) of anything that looks like a genuine mention rather than an unrelated result. Capture per mention: URL, source name/platform, published date, author/handle if public, and the text that mentions the brand.

Step 4 — Dedupe against the ledger

Compute a stable fingerprint per mention (the URL, or a hash of URL plus the first few hundred characters). Compare against memory/brand-monitor-log.md. Anything already logged is skipped — this run only classifies and reports mentions that are new.

Step 5 — Classify sentiment and notability

For every new mention, call the sentiment: positive, neutral, or negative, with a one-line reason. Flag as notable anything that is:

• Negative in sentiment.
• High reach (large follower count, high upvotes/comments/shares).
• From an influential source (press outlet, verified or large account).
• Factually wrong about the brand in a way that could spread.

Everything else is routine — still logged, but reported as a one-line rollup, not a full callout.

Step 6 — Draft a suggested response for flagged mentions

For each notable or negative mention, draft exactly one suggested response: calm, on-brand, addressing the specific concern raised. Label it clearly — "Suggested response (draft — needs human approval before sending)." Never post it, reply with it, or publish it anywhere. It exists only inside the Slack digest for a person to read, edit, and send themselves if they choose to.

Step 7 — Compose and post the digest

One message to {{slack_channel}} per run:

• A count of new mentions found and the sentiment breakdown.
• For each notable/negative mention: source link, a short excerpt, the sentiment and why it was flagged, and the suggested response draft.
• Routine new mentions: a one-line rollup (source + one-line gist), not a full callout each.
• A quiet run (no new mentions): a brief note, not a wall of "nothing found."

Step 8 — Update the ledger

Append today's mentions to memory/brand-monitor-log.md — fingerprint, URL, source, sentiment, notable flag, and whether it was reported — plus a dated run-log line summarizing what was posted.

</workflow> <ledger-format> Lives at `memory/brand-monitor-log.md`. Per mention: fingerprint (URL or URL+content hash), source/platform, date first seen, sentiment classification, notable flag (y/n), and whether it was reported in the digest. Below that, dated **Run log** entries: how many new mentions were found, the sentiment breakdown, and which ones got a suggested response draft. </ledger-format> <guardrails> - **Monitor and alert only.** The agent never replies, comments, or posts on any external platform — not the source of a mention, not anywhere else — under any circumstance. - **Draft, not send.** A suggested response is always a labeled draft inside the Slack digest, for a human to review, edit, and send themselves. The agent never sends it on its own. - **Read-only against the web.** Public search and fetch only — no login, no auth, no write of any kind to any external site. - **One output surface.** {{slack_channel}} is the only place this agent writes. It has no access to internal systems and takes no action beyond posting the digest. - **Brand terms are code.** Changes to the watch scope go through the brand terms file and a reviewed change request, not an ad hoc edit mid-run. - **Secrets scoped.** No credential is needed for public search and fetch today. If a future source requires auth, that credential is encrypted in the secrets manager and injected at runtime — scoped to this agent's grant. </guardrails> </skill>