competitor-brief

name	competitor-brief
description	Recurring competitor and market monitoring. Tracks a saved list of competitors, checks their sites and recent news each run, diffs against the last run, and reports only what materially changed — pricing, launches, funding, hiring, positioning. Use for "track competitors", "competitive intel", "what changed with <competitor>", managing the tracked list, or the weekly competitor brief.
compatibility	Designed for Vellum personal assistants
metadata	{"emoji":"🔭","vellum":{"category":"productivity","display-name":"Competitor Brief","includes":["admin-copilot-prefs"],"activation-hints":["User wants ongoing competitor or market monitoring","User asks what changed with a competitor recently","The weekly competitor brief schedule fires","User wants to add, remove, or list tracked competitors"],"avoid-when":["User wants a one-off deep research report unrelated to tracked competitors (use deep-research)","User wants a general web search with no competitor-tracking intent"]}}

Competitor Brief

Keep a small, durable picture of each tracked competitor and surface only what changed since last time. The value is signal, not a re-scrape: a good brief is short, dated, and sourced.

State lives in the admin-copilot plugin's storage and is read/written through the admin_copilot_prefs tool from the included admin-copilot-prefs skill, called via skill_execute — never edit files by hand:

• The tracked list: list_competitors, add_competitor, remove_competitor.
• Per-competitor last-run snapshots: get_competitor_snapshot, save_competitor_snapshot.

If skill_execute reports admin_copilot_prefs as unknown or not allowed, the admin-copilot-prefs skill has not projected into this session — the state actions will not become reachable by exploring, and searching the filesystem for the storage files will not find them (the path is internal). Try skill_load with skill: "admin-copilot-prefs" once; if it still fails, do not loop or hand-write state. Stop and report one line: "Competitor brief can't run yet — the admin-copilot state tool isn't loading; restart the assistant to fix it." Do not fabricate a brief from memory.

When the user is curating the list

• Add: admin_copilot_prefs add_competitor with name and, when known, domain, watch_urls (pricing / changelog / blog / careers pages worth checking), and keywords (topics to weight — e.g. "pricing", "enterprise", "funding").
• Remove / list: remove_competitor / list_competitors.
• Confirm the resulting list back to the user.

Producing a brief (per run)

1. Load state. list_competitors, and for each one get_competitor_snapshot to recall what you saw last time (sources + findings + lastRunAt).

2. Gather current signal for each competitor:

   • Fetch each watch_url with web_fetch; run a web_search for the competitor name plus its keywords and recent dates.
   • For a deeper periodic pass (e.g. monthly, or when the user asks for depth), escalate to the deep-research harness instead of light fetches. Keep the weekly run light and cheap.

3. Cheap change pre-filter. For each fetched page, hash the text and compare to the stored source hash to skip pages that did not change at all:

   # prints the sha256 of normalized stdin text
   echo "$PAGE_TEXT" | bun "$SKILL_DIR/scripts/page-hash.ts"
   

   A matching hash means the page is byte-stable — skip deep analysis of it. A new or changed hash means read it closely. The hash is only a pre-filter; you decide what is material.

4. Judge materiality. Compare new signal against the prior findings. Surface only genuine changes: new/changed pricing, product launches or major changelog entries, funding or M&A, notable hiring (esp. leadership), and positioning/messaging shifts. Ignore cosmetic edits and re-phrasings.

   Suggest a next move (don't take it). A finding that just sits there is half the value. For each material finding, decide whether it warrants a concrete response — and if so, prepare it, never transmit it:

   • A competitor pricing or positioning shift sales should know → a draft note to the sales/team channel (e.g. a slack draft), handed over for the user to send.
   • A launch or capability worth a closer look → propose a deeper deep-research pass, or a calendar hold to dig in.
   • Something worth keeping in the user's own context → a one-line remember.

   Only attach an action when there's a real one; most findings are just signal. No action ever leaves the building on its own — every note is a draft the user sends, consistent with the no-send rule below.

Empty weeks

If nothing material changed, still report — one line: "No material competitor changes this week (checked N competitors)." Silence reads as a broken job and erodes trust. Always save the refreshed snapshot even when findings are empty so hashes stay current.

Rules

• Only changes worth a busy person's attention. When in doubt, leave it out.
• Every finding cites a source URL. No source, no finding.
• This skill never sends email or messages on its own — delivery to slack/email goes through the dedicated channel skills, which carry their own confirmation.
• Keep weekly runs cheap; reserve deep-research for explicit deep passes.