research-competitors

Competitor research tells the marketer what to do next; it is not a list of ads. Craft: references/research-spec.md; worker briefs: its step skills (research-competitors-<step>); the report: references/report-format.md — next to this file.

The rule that matters most: free pass first, ask nothing, report once (../marketing-guide/research-pattern.md (in the marketing-guide skill) — read it before the first search). For a full teardown, research every dimension for every main rival, browser-only sources included, before the first message; never ask permission for free work or offer to pay for what a free read could get. An ask that names an area ("what ads are they running?") gets that area covered completely, news and launches first.

Before you start

Setup first: do the marketing-guide skill (../marketing-guide/SKILL.md) §1–2 (connect, brand, brain read) before anything else; the job line is "I research your competitors and save what I find through it." Here: a named brand not in the account → create it from its website (gooseworks:brand_setup_create) without asking and wait for its research (gooseworks:job_get); don't show a menu unrelated to research. Fallback brand read: gooseworks:brand_read (summary, kit, products, competitors, organic, learnings). Either path: gooseworks:competitor_read (+ suggestions), gooseworks:automation_list; workspace: research/competitors/ (last report, rival records), research/market/ (the newest market report and its "which market" decision), research/audience/ (buyer-groups.md and the latest audience report), plans/, reports/performance/. Brand facts: brain first, website second. Claims: only the read step of the claims-gate skill (../claims-gate/SKILL.md) now.
Focused output: for a battlecard, page tools/pixels, or LinkedIn engagers, select and finish the matching branch in references/focused-outputs.md after setup, then return. Reuse saved evidence; these branches replace the full teardown below.
Run memory: job competitor research or competitor watch; what this job hands the atom is in references/run-memory.md. Its "since last time" opens the chat and the report's story.
Never change the host's settings or permissions, and never seed a search or a worker with names or claims from memory — only what a source gave you.

Steps

Find the field yourself, before any worker starts. About 6–10 searches by the main run: news for our brand and the category (last 90 days), then "best ", " alternatives", " alternatives", " tools", and one AI answer. The saved roster is a starting list, never the answer. State the frame of reference in one line — the category we compete in (research-market's decision when it exists) and the other frames we could claim (e.g. "AI ad-creative tool" vs "AI agents for go-to-market teams"), because the frame decides which rivals matter. Then choose 3–6 main rivals — the ones buyers compare us with now, each with its reason — and 2–4 adjacent options (doing it yourself, an agency, the platforms' own tools). If the frame or the main rivals differ from the last run's, say so in one line with the reason ("Last run compared us with Jasper and Copy.ai; this run with Creatify, because…") — the frame never swings silently. A name you'd only watch goes in the brief competition line, not as a main rival. Anything from the last 30 days is a candidate timely opening.
Brief the workers — lean. Exactly one subagent per main rival with the research-competitors-rival-pass skill (../research-competitors-rival-pass/SKILL.md) (a new entrant you'd call closest to our buyer, or hitting us harder than a main rival, counts as a main rival — it gets its own worker, never just a name), one with the research-competitors-category-pass skill (../research-competitors-category-pass/SKILL.md), one with rival-pass on our own brand — no more, no workflow on top. Paste the short brief's full text with brand_id, our product summary, approved claims, the frame, a typical-buyer scenario for pricing ("a 10-person team making 40 ads a month"), the off-limits list and today's date; workers don't load skills or start their own subagents. Standard depth is the brief's cap (about 12 searches and 12 fetches each); a deep teardown — quarterly, or when the marketer asks — doubles it. A cheaper model is fine. Workers use web search, page fetch and the free server reads, and return everything, including the pages that need the browser. Edge cases — no browser, a stalled or stopped worker, facts from earlier research, a bad rival on the list, a report already saved today, a late read — follow the research-competitors-edge-cases skill (../research-competitors-edge-cases/SKILL.md).