Research supported advertising promises from a specific buyer's context, current product facts, customer language and observed ads. Turn them into static, video or copy ideas; match video ideas to actual recipes and save a product-scoped angle bank for write-video-ad-script. Product-led discovery can use verified facts without reviews. Reference persistence and reach are observations, not proof of performance.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Summary. Research advertising promises and match them to creatives the brand can make. Video research produces a structured bank the script writer can reuse, with stable angle ids, separate product facts and customer language, and template compatibility. Research scores rank test candidates; they do not predict conversions.
An ad is copy plus creative. This skill finds the angles real buyers respond to and turns each into an ad idea the user can make:
Core principle: combine audience relevance with a supported product promise and a feasible creative. Customer language, demonstrations, first-party performance and advertiser persistence provide different evidence. Long-running ads and variants are observations, not proof of profit. A product-led angle can start from verified facts without reviews; label missing evidence.
Do only the work the chosen output needs. Copy and static never search for videos or read video formats. Copy never looks at templates. Video reuses available reviews for language and insight; collect more only when the angle needs them. A video request may name a selected template or angle: preserve both.
scrapecreators-api,
competitor-ad-intelligence) and web search. Static and video ideas then describe the layout or
format in words instead of naming a library entry, and there is no hand-off step.Read it from the request:
| The user says | Output |
|---|---|
| static, image ads, banners, "what image ads" | static |
| video, UGC, reels, TikTok, "what video should I make" | video |
| copy, headlines, primary text, "what should our ads say" | copy |
| "ads" or "angles" with no hint | ask once: "Static ad ideas, video ad ideas, or just the copy?" |
The user can pick more than one (for example static and video); then run the union of their sources, once. Never widen the scope on your own.
| Source | static | video | copy |
|---|---|---|---|
| Brand context (1-0) | yes | yes | yes |
| The brand's own ads (1A) | yes | yes | yes |
| Competitor ads (1B) | yes, image ads first | yes, video ads first | yes, the copy |
| Customer voice: reviews + comments (1C) | yes | reuse reviews; comments on top posts | yes |
| Organic short-form reach (1D) | no | yes | no |
| Internal data (1E) | if provided | if provided | if provided |
| Reddit and forums (1F) | only if asked | only if asked | only if asked |
| Template library (2.5) | yes | no | no |
| Video format catalogue and full recipes (2.5) | no | yes | no |
Take what the request and the brand context already answer; ask only for what's missing, in one message:
Never invent a customer, a result or a claim: an angle may only promise what the brand's own facts support.
Run only the sections Phase 0B selected. ScrapeCreators calls go through scrapecreators-api
(GooseWorks: data_call_provider with provider scrapecreators, GET). Each is billed per call:
before the first one, say roughly how many you'll run and ask one yes/no. Free sources never need
asking. Large results are saved to a file by some hosts: request a full page and parse the saved
file rather than paging in small pieces.
With the GooseWorks MCP: brand_read (one brand in the org means use it and say which), with
sections summary, products, competitors, accounts and learnings (the reply is large: parse the
saved file); add kit when you need claims,
voice or photos. Note what it sells, who buys it, its claims and can't-say rules, its offers, which
products have clean photos (check the files: an SVG logo or a small thumbnail is not a clean
product photo), and whether it has screen recordings or footage. competitor_read lists tracked
competitors; with none, its suggestions view names likely ones. Say them in one line and carry on.
What the brand already runs decides what is new.
ads_template_read in competitor mode, one tracked competitor at a time
(filter by its source id). Rows are heavy; keep the source ad id, primary text, CTA, start and
end dates and ratio. Today most imported ads are images; their copy is still evidence.What buyers say in their own words. This is what keeps angles from being guesses.
The brand's own reviews: its product pages (review widgets often load by script, so a plain web fetch shows only a few; fetch the raw page HTML and read the review blocks and any "customers say" summary in it), its TikTok Shop reviews if it sells there (/v1/tiktok/shop/product/reviews), its Amazon listing (web search and fetch). Keep 4-5 star reviews for outcomes and proof, and 1-3 star reviews for objections to answer.
Competitors' bad reviews: 1-2 star reviews of the top 2 competitors' products, from the same kinds of sources. These suggest possible gaps to investigate. Claim a gap only when this product independently supports the promised difference.
Comments on what's working: most ads link to a landing page, not a post, so their comments usually can't be read. Read comments instead on:
Comments paths: TikTok /v1/tiktok/video/comments, Instagram /v2/instagram/post/comments, YouTube /v1/youtube/video/comments, Facebook /v1/facebook/post/comments. Look for questions, objections and "I switched because…".
Keep every quote verbatim with its link. Never paraphrase a quote into something the buyer didn't say.
What is getting earned reach right now.
Free first: the competitor dossiers (competitor_read with a slug) and social_inspiration_library
/ social_inspiration_search. These are often empty; then go to the paid searches.
Paid searches (one yes/no for the batch). Terms: the user's own first, then the category, the main problem it solves, each competitor's name.
| Platform | Path | Query | Notes |
|---|---|---|---|
| TikTok | /v1/tiktok/search/keyword | query, date_posted last-3-months, sort_by most-liked | About 2MB per call: parse it; keep url (tiktok.com/@author/video/id), author, followers, statistics.play_count, create_time, desc, commerce_info |
| Instagram Reels | /v2/instagram/reels/search | query, date_posted last-month | Google-indexed, best-effort, about 9 results a page; no follower count |
| YouTube Shorts | /v1/youtube/search | query, type shorts, nothing else | Adding uploadDate or sortBy with type shorts returns nothing. Rows have only url, title and views: for the 3-5 you'd cite, call /v1/youtube/video for channel, publishDate and isPaidPromotion. YouTube evidence is evergreen: cite its date |
| X | none | none | No X keyword search: use competitor_search_mentions with platform x, or /v1/twitter/user-tweets for a handle |
The full list is the official OpenAPI (docs.scrapecreators.com/openapi.json). If a call errors, read it there; never guess a path.
Keep vertical videos only, far above their account's usual views, from the last ~90 days.
Watch the 3-5 strongest (the watch skill, or social_inspiration_watch for saved posts). When
neither is available, read the transcript (YouTube /v1/youtube/video/transcript; otherwise the
caption and first spoken line) rather than guessing from the title.
Support tickets, NPS comments, sales call notes the user provides. Treat them as customer voice.
Only when the user asks, or for a product whose buyers mainly talk there (developer tools, some
software). Use reddit-post-finder or /v1/reddit/search. A thread is one person's view: it counts
as customer voice only when the same point also shows up in reviews or comments.
Reviews and comments are customer voice, a third kind of evidence. Label them as such.
Read buyer case and supported product role.
Before naming an angle, extract a specific current consumer task or decision and
explain why this exact product can help. Record the buyer's inputs, current approach,
constraint and desired output, plus the supported product action and useful role.
Keep source limits and authored scenarios explicit. Carry these as buyer_case and
product_role in the video handoff; a generic pain label or ingredient list is not
enough. Preserve human feedback separately from model ranking.
| Category | What to look for |
|---|---|
| Pain | Specific frustrations with the status quo or competitors |
| Outcome | Results buyers describe in their own words |
| Identity | How buyers describe themselves or want to be seen |
| Switching | Why people left a competitor |
| Proof | Outcomes, ratings or credentials buyers and the brand can back up |
| Contrast | Old way vs new way, them vs us |
| Objection | The doubt that stops a purchase, answered |
For each angle record: the one-sentence angle, relevant verbatim quotes with links when available, and the sourced product facts or observations that support it. Do not invent quotes to meet a quota or require a complaint for a product-led discovery angle. State which kinds of evidence back it (customer voice, product facts, organic reach, sustained ads), whether the brand already runs it (1A: measured winner, persistent, inactive, or never observed) and the competitor gap it exploits, if any.
Every output gets the words: a headline (8 words or fewer for static, 12 or fewer as a video hook) and 1-2 sentences of primary text, in the brand's voice.
Claims: only what the brand's own data says (certifications, guarantees, ingredients, offers), checked for the exact product the ad shows: two products of one brand can differ (one has no added sugar, the other does). Flag anything the brand should approve, such as review counts, health benefits, price comparisons or a named person, rather than writing it as fact. When the brand's sources disagree (two different numbers for the same claim), flag both.
The template library grows every week, so never rely on a remembered list of templates or layouts.
ads_template_read in query mode, with the brand id and that sentence as the
query. It searches the whole library by meaning and returns each match with its description,
search tags and a score, so templates added tomorrow are found the same way. Rank by score.
Run one search per angle; when the best similarity is under about 0.5, try a second phrasing.goose-ads as a
community ad id, a template as a template id.Read the live video format catalogue for the brand. Only listed formats can be made. When the brief already chooses a template, use it. For plausible candidates, fetch the full template recipe before calling the idea make-ready: a card description is not a production contract.
Match the angle to the recipe's story mechanism, speakers, timing, text slots, proof device, permitted visuals and required assets. Keep the quoted catalogue description for display, and record the actual template id and compatibility reason. A physical demo needs a real product; a screen demo needs usable recordings; a silent reveal needs a visual payoff. Do not turn all angles into testimonials. An unmet essential input makes the pair provisional or incompatible.
For dialogue directions, return an observed same-format reference alongside the angle. Reuse collected sources before retrieving more. A caption, format description or imagined beat map cannot establish human dialogue.
Street ads require a complete commercial interaction, with full visual timeline and spoken exchange inspected. Match physical/service/digital offering and the actual interaction (sampling, mic-only conversation, or challenge). Preserve source, observation method/limits, timestamped words and actions, visible setup, participation reason, hook, brand connection, payoff, and unseen recruitment as explicitly unknown. Label paraphrases and retain full transcript evidence. Partial hooks and editorial radio excerpts cannot fill this gap; seed summaries are source leads until inspected in full. The selected [[references::render-street-interview]] guide and writer's files reference define the record.
For podcasts, an appropriate editorial interview can provisionally teach turn structure, with commercial and delivery limitations recorded. Neither real performers nor an observed transcript establishes human authorship or conversion performance.
Carry audience, objective, offer, CTA and exact product through this match. When the angle or format is fixed, do not replace it to improve a ranking. Report a conflict and suggest a compatible execution. Never infer that a renderer can make a scene from a generic model name.
No template or format. Give 2-3 headline and primary-text variants per angle instead.
Adapt the pattern, never copy: take the shape, structure and angle, never another brand's words, claims, faces, footage or offer. Spread the list: no more than 3 ideas on one template or format, and no more than 3 on one angle (variants of one angle count toward that 3).
| Factor | Weight | Description |
|---|---|---|
| Evidence | 40% | How many kinds back it (customer voice, organic reach, sustained ads) and how strong each is |
| Gap | 20% | Never run by the brand, and weak or absent in competitors' ads |
| Fit | 20% | The brand can claim it with its own facts, and buyers like its own customers say it |
| Make-ready | 20% | The template or format fits and the brand has what it needs (copy: 100%) |
Reject unsupported product promises and impossible angle–template pairs before scoring. Score the remaining pairs out of 100 and rank as research priorities. These weights and caps are heuristics, not calibrated probabilities of performance. Keep a measured first-party winner separate from the research score. Do not pad the list when evidence is thin.
Print one table in the chat, best first, with up to 10-15 distinct, supported ideas. Every idea needs at least one link a tool actually returned; no link, no idea.
Static:
| # | Headline | Primary text | Angle | Template (why it fits) | Evidence | Needs | Score |
|---|
Video:
| # | Hook | Angle | Format (quoted card) | Evidence | Needs | Score |
|---|
Copy:
| # | Angle | Headline variants | Primary text | Evidence | Score |
|---|
In Evidence, label each link: customer voice, organic (with views vs usual), or paid (with days running and variants).
Under the table:
For video, read the video-handoff reference and save video-angle-bank.v1 with stable angle ids, brand and product scope, audience/objective/offer/CTA, sourced facts, verbatim quotes, observed references and compatible template ids. Save it with the workspace file tools in the brand and product's video-scripts area as angle-bank.json, beside customer-words.json. Also keep a copy in this run. If workspace writes fail, retain the run copy and pass it directly; say cross-session reuse is unavailable. Do not pretend a bank was persisted.
A human-readable angle-bank Markdown and ideas JSON remain useful for static and copy output. For video they may accompany the shared bank, but do not replace it. The script writer consumes the bank rather than trying to reconstruct research from a chat table.
Then ask: "Which ones should I make? Pick up to 5, or say 'the top 3'." (Copy: skip this.)
For the picked ideas, with no manual step in between:
goose-ads skill as the source and the brief.video_project_upsert (brand id, a short name from the
hook, format set to the template id, and no brief, since a brief makes a multi-concept batch).
Then make them one after another with goose-video-local; each idea is that project's brief.
Pass the complete selected idea and shared bank into goose-video-local, including the chosen
angle id, product id, template id, evidence and proof requirements. Its script step must fetch
write-video-ad-script and prepare a validated angle context; never pass only the hook. Keep raw
buyer records outside remixable review data. The local-template path needs a terminal agent;
hosted/server formats have their own runtime and require the same strategy contract there.Up to 5 per request; offer the rest after. Each paid step of making an ad is approved before it runs.
brand_read, competitor_read, competitor_search_mentions,
ads_template_read, social_inspiration_library, social_inspiration_search,
data_call_provider, video_catalog_list, video_project_upsertscrapecreators-api: ad libraries, comments, reviews, social searchcompetitor-ad-intelligence: deeper teardown of one competitor's ads, when askedcomment-mining: structured mining of a long comment threadreddit-post-finder: only when Reddit is asked for (1F)まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Analyze ad campaign performance data (Google, Meta, LinkedIn) to identify what's working, what's wasting budget, and specific cut/scale/test recommendations. Runs statistical analysis, funnel diagnostics, and multi-channel budget reallocation with specific dollar-amount shift recommendations and scenario modeling.
日本語の概要は準備中です。原文の説明を表示しています。
For paid lead-gen and participant-recruitment ads, replaces vanity CPA with true CAC per qualified lead by joining ad-platform data with downstream funnel events, surfaces tracking gaps, and classifies every creative into Scale / Keep / Investigate / Cut.
日本語の概要は準備中です。原文の説明を表示しています。
The intake stage of a Meta ad harness. Researches the brand first, states what it assumes, and asks only what it cannot infer and what would change money, claims or legal exposure if guessed wrong — at most four questions a turn, in plain business language. Writes the brand's stable facts (ads/brand.md), the campaign brief, and a log of what was asked vs assumed. Use it when someone wants to run Meta ads and the brand's ad facts are not yet recorded.
日本語の概要は準備中です。原文の説明を表示しています。
Analyze the message match between your ads and landing pages. Checks if the promise in the ad copy carries through to the landing page headline, body, and CTA. Flags disconnects that kill conversion rates. Works with Google, Meta, and LinkedIn ads.
日本語の概要は準備中です。原文の説明を表示しています。
Turn a brand's facts, a campaign brief and the product into a one-minute Meta ad strategy with its reasoning — goal, audience, objective and placements, budget split, how many ads and how different they must be (generate 3x so the user can choose), the test frame, the success metric, and safety rules for the user to accept. Decides what launch-meta-ad-campaign then executes, and re-reads the strategy after a deep check to show which assumption changed. Use it after intake and before any creative is generated or any campaign is built.
日本語の概要は準備中です。原文の説明を表示しています。
Check and improve your brand's visibility across AI search engines (ChatGPT, Perplexity, Gemini, Grok, Claude, DeepSeek). Set up tracking, run visibility analyses, audit your website for AI readability, and get actionable recommendations. Uses the npx goose-aeo@latest CLI.
日本語の概要は準備中です。原文の説明を表示しています。