competitor-diff

<skill name="competitor-diff"> <overview> Watch competitor sites, changelogs, and pricing pages without turning into an alert firehose. A daily cron re-prompts a persistent session that fetches each page on the watch list, diffs it against the stored snapshot from the previous run, filters out cosmetic noise, and posts a short summary of what actually changed to {{slack_channel}}. On a day nothing moved, the channel stays quiet.

Proactive and schedule-driven; read-only against the public web, with a single Slack channel as the only output. </overview>

<when-to-load> - The daily cron fires the competitor-watch run. - A human asks for a manual check of the watch list. - The watch list itself needs updating (a competitor added, dropped, or a tracked page's URL changed). </when-to-load> <workflow>

Step 0 — Orient and resume

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

Read the last snapshot recorded for each tracked page. If this is the first run ever, there is nothing to diff against — fetch, store the snapshot, and skip reporting (nothing to compare yet).

Step 1 — Load the watch list

The watch list lives in {{watch_list}} (competitor name → URL, one per line, e.g. Acme — pricing: https://acme.com/pricing). Treat it as the source of truth for exactly which pages to fetch this run — no more, no less.

Step 2 — Fetch each tracked page

Fetch every URL in the watch list over plain HTTP(S) GET. No auth, no credentials — these are public pages.

curl -sL --max-time 20 "$URL"

If a fetch fails (timeout, 404, page moved), note it in the ledger as a blocker for next run rather than failing the whole sweep.

Step 3 — Normalize before comparing

Strip what varies without meaning anything: nav, footer, ads, timestamps, session-specific query params, whitespace. Reduce each page to its meaningful text content (main content area, changelog entries, pricing table rows) before computing a diff or hash.

Step 4 — Diff against the last snapshot

Compare the normalized content to the snapshot stored in memory/competitor-watch-log.md for that URL. A byte-identical normalized page is "no change" — move on without reporting it.

Step 5 — Filter cosmetic from meaningful

Not every text diff is worth a Slack message. Report:

• A pricing tier, price point, or plan name that changed.
• A new changelog entry or shipped feature.
• A rewritten headline, tagline, or positioning statement on a landing page.

Drop:

• Typo fixes, formatting, reflowed layout with the same words.
• Date/time stamps, view counts, or other page furniture.
• Reordering of existing content with no wording change.

When in doubt whether a diff is meaningful, err toward including it — under-reporting defeats the point of the watch.

Step 6 — Compose and post the summary

One message to {{slack_channel}} per run:

• Something changed: a short bullet per meaningful change — competitor, what changed, and the page it happened on. Keep it scannable, not a raw diff dump.
• Nothing changed: a single brief line ("No meaningful competitor changes today.") or skip the post entirely per the project's preference — never post per-page "no change" noise.

Step 7 — Update the ledger

Update memory/competitor-watch-log.md with today's normalized snapshot for every fetched page (so the next run has something to diff against) and a dated log line of what was reported or skipped.

</workflow> <ledger-format> Lives at `memory/competitor-watch-log.md`. Every run updates, per tracked URL: the competitor name, the URL, the normalized snapshot (or a hash of it) from this run, and the timestamp it was fetched. Below that, dated **Run log** entries with what was reported to {{slack_channel}} (or "quiet — no meaningful changes") and any fetch failures to retry next time. </ledger-format> <guardrails> - **Read-only against the web.** GET requests to public pages only — no login, no credentials, 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 summary. - **No noise.** Cosmetic edits are filtered before they ever reach a human; only pricing, feature, and messaging changes are worth a message. - **Watch list is code.** Changes to which competitors/pages are tracked go through the watch-list file and a reviewed change request, not an ad hoc edit mid-run. - **Secrets scoped.** If a future watch-list page needs auth, that credential is encrypted in the secrets manager and injected at runtime — scoped to this agent's grant. (None are needed today.) </guardrails> </skill>