Agency Brand Scoping

Agency brand scoping

A fast discovery pass for a new client: gather brand signals (site, deck, competitors, audience, tone), lock them into a reusable brand-system.json file, and produce 5 on-direction visual explorations to validate with the client before any production spend.

URL / deck in → brand-system.json + 5 direction variations out. One hour, under $1 in generation cost, reusable across the rest of the engagement.

When to Use

• Responding to an RFP and need 5 distinct visual directions for the first review
• Net-new client pitch — you've read the brief, now you need on-brand sketches
• Kickoff phase of a signed engagement, before production generations start
• Scoping a rebrand: grab the existing brand, propose 5 evolution paths
• Any moment where "what does this brand look like in AI generation" is the open question

Do not use for finished campaign assets — this is discovery only. Lock the system here, then run agency-pitch-mockups or agency-multi-brand-pack for deliverables.

Prerequisites

Ask the user (batch in one message):

1. Client name + slug — used for folder + manifest tags (e.g. acme-fintech)
2. Brand references — URL, existing deck, Figma file, or "no brand yet, we're defining it"
3. Competitors / comparable brands — 2-3 names; informs what NOT to look like
4. Audience + tone — who buys, what feeling ("premium + restrained" vs "bold + irreverent")
5. Deliverable type the scope is for — pitch deck, campaign, launch film, product shoot (informs aspect ratios)
6. Confidentiality — is this NDA? If yes, never name the client in public Drive folders or prompts

If the user gives a URL or deck, read/fetch it first and extract palette, typography impression, imagery style, and tone words. Propose the brand-system.json back for confirmation before generating.

How to Run

1. INGEST     → read URL / deck / Figma, extract signals
2. DRAFT      → propose brand-system.json (palette, type-feel, imagery, tone, do-nots)
3. CONFIRM    → user locks the system; save to clients//brand-system.json
4. ESTIMATE   → gen-ai pricing on the 5-direction batch (/brand.md for all future work

Rules:

• Never skip step 2. Without brand-system.json, every future generation for this client re-litigates brand from scratch.
• Always estimate before generating. 5 variants at recraftv4 ≈ 10 credits, but confirm in-session.
• Keep pitch outputs watermark-free only if the client has signed. For cold pitches, add "obvious mockup" language in prompt.
• Never mix clients in a single Drive folder or manifest. Namespace everything under clients//.

Quick Reference

{
  "client_slug": "acme-fintech",
  "brand_system": "clients/acme-fintech/brand-system.json",
  "defaults": { "model": "recraftv4", "aspectRatio": "16:9" },
  "variants": {
    "direction": ["editorial", "bold", "minimal", "playful", "cinematic"]
  },
  "jobs": [
    { "id": "{direction}", "prompt": " — {direction} direction" }
  ]
}

Lock the winning direction back into clients//brand.md — a 10-line markdown file Claude + downstream batches can include in prompts.

Quick Reference

| Sub-task | Model | Why | |---|---|---| | Direction exploration (default) | recraftv4 | Design-forward, honors palette + type feel, cheap | | Photo-led brands (fashion, hospitality) | flux-2-pro | Photoreal, better lighting, for brands that live in photography | | Typography-centric directions (editorial posters, quote slides) | ideogram-v3 | Only model that renders readable headlines reliably | | Quick cheap drafts when testing 10+ directions | gemini-3.1-flash-image or gemini-3.1-flash-image | ~1 credit each; use for throwaway iteration |

Check live IDs with gen-ai models --mode image before committing — names shift.

Procedure

• Always scope brand before generating production work. One hour of scoping saves a week of misaligned deliverables.
• Descriptor-per-direction beats vague prompts. "Editorial" + specific cues > "something clean and modern".
• Never leak other clients' assets into a new client's prompt. No -i from a different client folder, ever.
• Save brand-system.json to the repo, not Drive. It's versioned alongside code and MRs.
• Propose 5, not 10. Client decision fatigue is real; 5 distinct directions forces a real pick.
• For NDA clients, don't name them in prompts. Use a generic descriptor ("a premium fintech for seed-stage founders") — prompts are logged.
• Build a reusable prompt library per client at clients//prompts/ — hero, tile, social, OG templates that all reference the locked direction.

Pitfalls

• Generating before locking brand-system.json → 5 gorgeous directions that don't fit the brand
• Brand cross-contamination — reusing a prior client's brand.md "because it looked nice" — your competitive moat is per-client rigor, don't blow it
• Descriptor overlap — "bold" and "cinematic" blur if prompts aren't distinct enough; each direction needs a clearly different visual hypothesis
• Over-polished scoping — if the scoping output looks like a finished ad, the client will ask "why not use this?" and you've skipped production
• Missing competitor check — landing on a direction that looks identical to the client's biggest rival
• Public Drive folder for an NDA client — always use --drive-folder "internal-$CLIENT" or a private workspace

Verification

Run gen-ai whoami to confirm authentication, then re-run the failed command with --debug.

Step 1: Ingest + propose brand-system.json

Pull signals from whatever the client gave you. If it's a URL, fetch it. If it's a deck, read it. Summarize into:

{
  "client_slug": "acme-fintech",
  "brand": {
    "palette": { "primary": "#0B1F3A", "accent": "#00D4A3", "neutral": "#F5F3EE" },
    "typography_feel": "geometric sans, tight tracking, editorial weights",
    "imagery_direction": "abstract finance, muted photography, no stock clichés",
    "tone_words": ["trustworthy", "calm", "precise"],
    "do_not": ["cartoon illustration", "neon gradients", "tech-bro stock photos"]
  },
  "competitors": ["wise.com", "mercury.com"],
  "audience": "seed-stage founders, CFO buyers",
  "confidentiality": "NDA"
}

Save to clients//brand-system.json. This file is the single source of truth referenced by every downstream skill.

Step 2: Generate 5 directions

Each direction gets a single-word descriptor. Standard starter set: editorial, bold, minimal, playful, cinematic. Swap any that clashes with the brand's tone words (e.g. drop "playful" for a private bank).

CLIENT="acme-fintech"
OUT="clients/$CLIENT/scoping"
mkdir -p "$OUT"

# Check per-call pricing first, then multiply by 5 directions.
gen-ai pricing recraftv4

# Run as 5 discrete jobs so each can get its own direction prompt
cat > /tmp/$CLIENT-scoping.json /scoping/0X-.webp` + `results.json`. Present the 5 side-by-side to the client.

## Cost & time

| Phase | Typical spend | Typical time |
|---|---|---|
| Scoping (this skill, 5 directions, `recraftv4`) | ~$0.50–$1 | 15-30 min incl. review |
| Production (after direction lock) | $5–$50 per asset set | Per deliverable |

Keep scoping under $1. If a client won't approve a direction after 5, the brief is broken — don't burn credits, push back on the brief.

## See also

- `workflows/agency-pitch-mockups/` — once direction is locked, produce pitch assets
- `workflows/agency-multi-brand-pack/` — run scoping-plus-production across many retainer clients
- `workflows/agency-client-handoff/` — package final work at engagement end
- `gen-ai-workflows.md` — general multi-step patterns
- `gen-ai-batch.md` — manifest shapes and concurrency tuning

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [PicsArt](https://github.com/PicsArt)
- **Source:** [PicsArt/gen-ai-skills](https://github.com/PicsArt/gen-ai-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.