gooseworks-ai/goose-skills / skills/marketing/research-market
research-market
VerifiedMarket research and brand watch: which market the brand is in, how big the reachable part is, what's changing, how buyers buy and what they pay, what people are saying now, and what it means for positioning. Use when the marketer asks for market research, "how big is my market", "what's changing in my category", "what do others charge", "what are people saying about us", "is there demand for <idea>", or "watch my brand / tell me if something blows up" (alerts only, risk first). Also writes positioning choices and finished messages ("position my product", "write our positioning"). Researches first, asks only decisions at the end, and saves the output to the brand. Not for who the buyers are and their own words — use research-audience; not for a deep look at rivals — use research-competitors.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 10.6 KB |
| references/idea-validation.md | markdown | 2.9 KB |
| references/listening-craft.md | markdown | 11.6 KB |
| references/positioning.md | markdown | 5.9 KB |
| references/report-format.md | markdown | 8.5 KB |
| references/run-memory.md | markdown | 9.4 KB |
| skill.meta.json | config | 4.9 KB |
| tests/examples/2026-10-22/positioning.md | markdown | 8.0 KB |
| tests/fixtures/positioning-evidence.md | markdown | 3.2 KB |
| tests/test-by-hand.md | markdown | 6.7 KB |
tests/test-by-hand.md
View on GitHub# Test by hand — research-market
## The test (rebuilt 18 Oct after the senior-marketer review)
Use the GooseWorks server the test folder names (staging). Asks, exactly as the founder typed them: **"i want to do the market research for gooseworks"**; later, **"What are people saying about me and my category this week?"**; and **"watch my brand"**.
Pass, all without being asked:
1. **No question before the report** (beyond picking or creating the brand). No change to the host's settings.
2. **It decides which market** — 2–3 candidate categories, the one buyers use, a recommendation — and **sizes it top-down and bottom-up** (likely buyers × price per buyer group, from research-audience's buyer groups, or provisional groups with "run audience research" as a decision).
3. It covers change (dated, primary-checked or marked), buying, pricing (8+ products with units), attention, demand and what people are saying; rivals stay in one short section; it never builds its own buyer groups or quote bank.
4. **It reads the sibling reports** and states any disagreement; it ends the story with **what this means for positioning**.
5. **The report passes write-report §3** (check its list: the top says what was looked at and the findings before the first move). Also: a baselines table; "For you" only where a finding matters; a local file with a clickable link plus the brand copy.
6. **The closing message:** three lines, at most five decisions, a repeat only with a reason (monthly pulse plus quarterly report; brand watch weekly or daily).
7. **"Watch my brand"** runs the listening alone, risk first, and says nothing beyond one line when nothing crosses a threshold; an alert has what, reach, separate people, what we know, who handles it, and a hold list for serious risks.
Fail (20 Oct hand test): "Do this first" before the reader knows what was researched; a same-week sibling report with a different first move or lead buyer, not named or settled; a report delivered over its word budget.
Fail: questions before the report; a report about rivals' complaints instead of the market; no market decision or no bottom-up size; buyer groups invented here that contradict research-audience's buyer groups without saying so; a report saved only on the server; ids or jargon; a brand watch that sends a full report when nothing happened.
**The 18 Oct hand test (before this rebuild)** turned into rival research, lost the brand's own Product Hunt comments when a worker stalled on a browser prompt for 52 minutes, used the website's prices over the brain's, and produced a report full of ids and agent terms.
## Staging run — 2026-10-16 (GooseWorks staging, brand "Clearday", acne-safe skincare)
Ran every free step up to the first paid call. Nothing paid was called.
| Step | Call | What came back |
|---|---|---|
| Connection | `brand_read` (list) | 8 brands; Clearday `0e5b2fb0…`, coworker `9d554b73…` |
| Brain read | `brand_read` sections summary, competitors, learnings, onboarding, accounts, sources | 1 learning (a video-script "dont"); onboarding profile mostly empty (role, ad spend, goals missing); saved sources = organic_social (own IG/TikTok + 5 rivals), meta_ads (own + 5 rivals), google_ads (own). **No review source** (no Trustpilot) saved. |
| Evidence | `knowledge_search` "customer reviews pains objections what buyers say" | Only 2 hits: an AI-generated "what customers are saying" box from a Google ad page (context, not quotable) and a site metric "4.9 out of 5 stars, 234 reviews". → The bank must be built from public sources; the brain has almost no verbatim customer evidence. |
| Competitors | `competitor_read` + `view: "suggestions"` | 5 tracked (Paula's Choice, The Ordinary, CLEARSTEM, CeraVe, Prequel) with dossiers; 5 suggestions (Skinfix, Naturium, Sofie Pavitt Face CONFIRMED pages; La Roche-Posay, Topicals UNSURE). |
| Demand proxy | `ads_template_read` mode competitor | Rival ads with `ad_started_at` (e.g. Paula's Choice "Pick 3 and save" from 2026-08-24) — usable for the ad-rhythm proxy. |
| Previous run | `file_list` coworker workspace | Only `agent-config/` — first run, no previous bank. |
| Write path | `file_write` → `media_upload` (scope brand, kind document, source workspace) → `knowledge_import` → `job_get` knowledge_ingest → `knowledge_search` document | Tested on the QA brand "MCP Video QA Alpha" with a one-row test bank: import `ready` in ~10 s, 1 chunk, found by search; free. `media_upload` with source `bytes` **rejects markdown** (images/audio/video only) → the atom uses the workspace source. Deleting the media asset (`media_update` delete) removed its knowledge too; test file and media cleaned up. |
**First paid step (not run):** the pull plan would have been — Reddit `/v1/reddit/search` for "Clearday skincare", "acne safe moisturizer", "hypochlorous spray" (1 page each); Reddit `/v1/reddit/post/comments` on the top 5 threads; TikTok `/v1/tiktok/search/keyword` "Clearday skincare review" + `/v1/tiktok/video/comments` on the top 5; competitor product-page reviews by browser (free). No ScrapeCreators price on record → the run would ask one yes for a single probe (the Reddit search), keep its results, show its real charge, then ask for the stage A total (and later a stage B total for comments).
## 22 October — positioning acceptance (GOOSE-4058)
**Input:** `fixtures/positioning-evidence.md`, fictional supplied product/competitor/buyer evidence and exact product/price approval, no selected option. Test date: 22 October. Setup/approval state is simulated; no live read/write, publication, browsing or spend.
**Ask:** “Write positioning options and finished messages for Papertrail.” Follow references/positioning.md using the supplied evidence rather than starting a market-size report.
**Actual local output:** `examples/2026-10-22/positioning.md` (1,143 words). Inspected two different buyer/problem choices, attribute→value→proof, counterargument, current-position conflict, finished statement/pitch/hero/subhead/CTA/three supports/boilerplate/buyer adaptations, unsupported-copy table with safe alternatives, and a test with observable pass/fail. Current positioning is not overwritten; neither a selected brief nor a paid test is fabricated. All local evidence links resolve. Skill validation passed (351 skills at that point); no owned-file installed-layout link error.
**Next host checks:** no sibling reports → fill only material gaps and label provisional; owner chooses option 2 → only that scope's current brief changes, hypothesis status remains; contradictory/unapproved time-saved line → excluded from usable copy; second same-day run → preserve history with suffix; next run → read actual prior test result, not infer success from a file. Live brand save, current-brief update and subsequent run-memory retrieval remain untested.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |