Brand Research Agent Skill

brand-research

Purpose

Given a brand (name and/or URL) and a product, produce the full creative-prep package for

it: research the product, analyze the brand's running ads, measure their editorial DNA,

source logos and reference photos, generate brand-anchored product + lifestyle imagery, and

write the brand-context documents a downstream ad/video pipeline consumes.

The output is a brand-context pack — a self-contained set of artifacts:

• brand-summary, visual-identity, competitors, audience — the core brand context.
• existing-ads — what the brand's running ads reveal that web research misses.
• brand-grammar — the brand's editorial DNA (archetype, pacing, caption style).
• an asset manifest cataloging every sourced + generated asset with its kind, name, and

usage note, plus the binary assets themselves (logos, reference photos, generated stills).

• a concept brief of brand-level ad concept seeds.

Everything is written into a single brand-pack directory under output_dir.

Inputs

• brand (required) — the brand, used as the pack's folder name, e.g. amex, liquid-death.
• product (required) — the specific product / SKU / offer to research, e.g. "Platinum

Card", "Sparkling Water". Disambiguates brands with many SKUs.

• brand_url (optional) — canonical homepage. Strongly recommended to avoid wrong-entity

confusion (e.g. Apple band vs. Apple Inc.).

• output_dir (optional) — directory to write the brand pack into. Defaults to ./<brand>/.

Resolve the location from this input; never hardcode a path.

• max_existing_ads (optional, default 10) — cap on how many of the brand's running Meta

ads to pull for analysis.

• brand_video_urls (optional) — the brand's own video URLs (launch films, demos). If

provided, run build-brand-clip-library afterward to cut them into a reusable clip library.

• concept_count (optional, default 6–10) — number of social-ad concepts to draft.
• skip_generation (optional, default false) — skip image generation and ship research +

brief only (useful when no image budget is available).

Composed Atoms

• source-company-existing-ads — download the brand's running ads from the **Meta Ad

Library** (via the Apify FB Ad Library scraper) into a raw/ folder with provenance.

• rename-and-index-ads — watch each downloaded ad, semantically rename it, and write an

INDEX.md (per-ad strategy + cross-ad patterns).

• analyze-reference-grammar — measure each ad's editorial DNA (cut points, pacing curve,

archetype, audio mode) into a per-ad grammar-profile.json.

• source-brand-assets — scrape logos + reference hero photos from the brand's site / press kit.
• understand-brand-assets — distill web research + reference photos into the visual-identity

content (colors, typography, photography style).

• analyze-ad-hooks — extract recurring hooks/motifs from the downloaded ads.
• generate-ad-concepts — produce the concept list for the concept brief.
• create-product-images-higgsfield-product-photoshoot — 4–6 hero/end-card product stills

(Higgsfield product-photoshoot on gpt_image_2).

• create-product-images-nanobanana — 8–12 vertical 9:16 lifestyle stills (Nano Banana Pro).
• External tools: a video-watching capability (frame extraction + transcription — e.g.

yt-dlp + ffmpeg + Whisper), and web search + fetch.

Workflow

• Disambiguate. Confirm brand + product resolves to one entity. If brand_url is

missing and the name is ambiguous, stop and ask.

• Scaffold the brand pack under <output_dir> — a brand-research/ folder for the

markdown docs, a brand-assets/ folder for the asset manifest + binaries (logos/,

reference-photos/, generated-product-shots/, generated-lifestyle/, songs/), and an

existing-ads/ folder (raw/ + renamed copies + a grammar/ subfolder). Only create

subfolders that will be populated.

• Web research via web search + fetch. Priority order: brand site → trade press

(Adweek/AdAge/Campaign) → reputable category reviewers. Capture: product overview,

mechanics/pricing, benefits, target audience, current named campaigns with dates, core

positioning, voice/tone. Record every URL with its access date.

• Pull and study the brand's running ads. Mandatory — research without watching real

ads misses how the product is actually shown and talked about.

• Run source-company-existing-ads (max_ads=max_existing_ads) to pull the brand's Meta

Ad Library ads into the existing-ads/raw/ folder. Manual fallback per that atom's docs.

• Run rename-and-index-ads to produce semantically named copies + an INDEX.md (per-ad

strategy + cross-ad patterns synthesis).

• Product deep-dive pass. Re-watch (or reuse the frame grids) specifically to extract

product mechanics the website doesn't show: how the product is held, used, applied,

opened, paired; in-app UI flows that appear on-screen; physical form factors and

packaging; claims/proof the brand leans on; demographics and contexts of the people

shown; objections the ads pre-empt. These feed existing-ads.md in step 7.

• Editorial grammar pass. Run analyze-reference-grammar on each renamed ad. Each run

emits a grammar-profile.json carrying the ad's archetype match + confidence,

cuts_per_10s[], mean shot length, payoff-hold ratio, audio mode, and aspect. These

per-ad profiles are the input to brand-grammar.md in step 7 — the brand's editorial

DNA, so a from-scratch ad can inherit the brand's cut rhythm and archetype defaults.

• If no live ads are found, write the "No live Meta ads found" stub in BOTH

existing-ads.md AND brand-grammar.md ("No live Meta ads found as of <date> — grammar

defaults will be picked at design-brief time") and continue.

• Source brand assets via source-brand-assets:
• Logos: brand press kit or Wikipedia SVG → brand-assets/logos/.
• Reference photos: 2–4 high-quality third-party shots of the product/hero →

brand-assets/reference-photos/. Mark "not licensed for redistribution" in each

manifest entry's description.

• Songs: if existing ads exist, extract the audio bed of one ad as a tone reference →

brand-assets/songs/.

• Generate brand-anchored imagery (skip entirely if skip_generation=true):
• 4–6 product-photoshoot stills via create-product-images-higgsfield-product-photoshoot,

grounded on the strongest reference photo so the SKU stays consistent →

brand-assets/generated-product-shots/.

• 8–12 lifestyle stills via create-product-images-nanobanana (Nano Banana Pro), 2k

vertical 9:16, same reference grounding → brand-assets/generated-lifestyle/.

• Write the asset manifest — a catalog with one entry per binary asset (every logo,

reference photo, generated still, song). Each entry records, at minimum, a stable id, the

asset's path (relative to the brand pack), a kind

(`logo | wordmark | product_photo | lifestyle | video_ref | style_ref | ui_ref | song |

asset), a short name to search by, and a description` of how/when to use it plus any

usage constraint (e.g. "not licensed for redistribution" on scraped photos). A vague

description defeats the file's purpose. Write it as brand-assets/manifest.json.

• Write the brand-research docs. Use these exact section headers, filled from

research — never leave a placeholder marker behind:

• brand-summary.md — What the company sells, Who they sell to, `Why people buy

(jobs-to-be-done), Brand voice in three words, What to never say`.

• visual-identity.md — Primary colors (hex), Typography, Logo usage rules,

Photography style, Off-limits styles.

• competitors.md — ## Direct (each competitor: one-line positioning, pricing tier,

and how <brand> wins / loses vs them) and ## Reference creative (links / vibes to

emulate or avoid).

• audience.md — Primary persona, Where they spend time online, `Objections they

raise, Proof points that land`. Include 3–4 distinct ICP segments and verbatim audience

phrasing (Reddit/forums) — these become VO seeds downstream.

• existing-ads.md — narrative synthesis of what the brand's running ads (step 4)

reveal that web research misses. Headers: Ads watched (count + date range + link to

existing-ads/INDEX.md), How the product actually shows up on screen, `Recurring hooks

and angles, Claims and proof the brand consistently leans on, Who's shown using it`,

Objections the ads pre-empt, Voice & caption treatment, Implications for new ads.

INDEX.md stays the per-ad catalog; this file is the synthesized read.

• brand-grammar.md — the editorial DNA synthesized from the per-ad

grammar-profile.json files. Headers: Dominant archetype (which creator-grammar

archetype the brand favors — e.g. creator-talking-head, vo-product-demo,

founder-monologue — with the per-ad split), Pacing curve (mean cuts_per_10s across

ads, range, payoff-hold use), Audio mode (music-only / vo+music / speech-only mix),

Caption family (burned karaoke / static lower-thirds / on-screen text bursts / none),

Hook construction (first 1.5s pattern), Defaults for new ads (recommended archetype +

cuts_per_10s target + caption preset a from-scratch ad should inherit). Human-readable

seeding, not a machine contract — pick the archetype from a small fixed vocabulary you

define up front and reuse across brands.

• asset-urls.md — every sourced URL with its access date (the provenance trail).
• ui-references.md — ONLY if the product has notable in-app/product UI worth

recreating; catalog the key screens. Omit the file entirely otherwise.

• Write concept-brief.md with sections: Observed patterns from existing ads (only if

step 4 ran), Strategic foundation, Concept ideas (concept_count, each with hook + format

+ 15s/30s beat-by-beat + why-it-works + the KPI it serves), Production notes, Open

questions. This is supplementary brand-level seeding.

• Write brand-research/video-research.json — a machine-readable companion that mirrors

the deep-research findings from existing-ads.md + brand-grammar.md so a host can ingest

them without parsing prose. Shape: `{ competitors:[{name,relationship,notes}],

existingAds:{count,source,recurringHooks[],recurringClaims[],talentProfile,

objectionsPreempted[]}, grammar:{dominantArchetype,cutsPer10s,audioMode,captionFamily,

hookConstruction}, hooks:[{line,archetype,sourceAds[]}], ctas:[{line,intent,sourceAds[]}] }`.

Omit fields you don't have; write the file only when ≥1 ad was analyzed (skip on the "no

live ads" path). The markdown docs stay the human-readable source; this is their structured

echo.

Decision Rules

• Refuse to proceed without a disambiguated brand+product. Don't guess between two entities

sharing a name.

• If a generated product image trips a provider safety flag, fall back to the alternate model

(gpt_image_2 → nano_banana_2).

• All generated imagery must use the same reference photo so the SKU is consistent across the

asset library.

• Never claim licensed rights to scraped reference photos. Always note "not licensed for

redistribution" in that asset's description in the manifest.

• If the image-generation budget is too low, emit a warning, skip generation, and still ship

the research docs + concept brief.

• If the existing-ads pull returns zero live ads (or all downloads fail), do not fabricate ad

observations — stub existing-ads.md per step 4 and omit the "Observed patterns" section of

the concept brief.

• This skill produces research, sourced assets, generated imagery, and a concept brief. It

does NOT clip the brand's own videos. When the brand has its own footage worth reusing (or

brand_video_urls is provided), run the sibling molecule build-brand-clip-library.

Output

The brand-context pack:

• brand-research/brand-summary.md
• brand-research/visual-identity.md
• brand-research/competitors.md
• brand-research/audience.md
• brand-research/existing-ads.md
• brand-research/brand-grammar.md
• brand-research/asset-urls.md
• brand-research/video-research.json (structured echo of existing-ads + brand-grammar; only when ≥1 ad was analyzed)
• brand-research/ui-references.md (only when the product has notable UI)
• brand-assets/ — the asset manifest + populated logos/, reference-photos/, and

(unless skipped) generated-product-shots/, generated-lifestyle/, songs/

• existing-ads/ — raw/ originals, semantically renamed copies, INDEX.md, and

grammar/<slug>/grammar-profile.json per ad

• concept-brief.md

Quality Checks

• All six required brand-research/*.md files (brand-summary, visual-identity,

competitors, audience, existing-ads, brand-grammar) exist with **no remaining

placeholder markers** and use the exact section headers above.

• existing-ads.md and brand-grammar.md either reference a populated INDEX.md + per-ad

grammar-profile.json files (≥1 ad watched) or carry the explicit "No live Meta ads found"

stub — never silently empty.

• brand-grammar.md names a Dominant archetype that maps to one of the fixed creator-grammar

archetypes and gives concrete numeric cuts_per_10s defaults (not "fast" / "snappy" prose).

• existing-ads/ contains raw/ originals, renamed copies, an INDEX.md whose per-ad blocks

were filled by watching the files (not guessed from filenames), and per-ad

grammar-profile.json.

• asset-urls.md cites real, dated sources for every research claim and sourced asset.
• The asset manifest parses, has one entry per binary asset on disk — every path is relative

to the brand pack and resolves to a real file; every kind is in the allowed enum; no entry

is missing name/description.

• The concept brief references specific moments from each analyzed existing ad (when step 4 ran).
• Generated imagery is visibly the same SKU end-to-end (same colorway, finish, branding).

Failure Modes

• Brand name collides with another entity and brand_url was not provided → refuse.
• Brand press kit is unavailable and no acceptable third-party reference photos exist → flag

and stop before image generation.

• Image-generation budget too low → warn, skip step 6 cleanly, still ship the research +

concept brief.

• The existing-ads pull is blocked by the Meta Ad Library (rate limit / scraper auth) → fall

back to the manual browser/curl path documented in that atom; if still empty, write the "no

live ads" stub in existing-ads.md and continue rather than halting.

• Individual downloaded ad files are unreadable → log each failure in INDEX.md, skip that

file, continue.