matteotitta Ad Creative — AI ad agency in Claude Code
Generate a full multi-format ad pack (static + animated + UGC, LinkedIn + Meta) from one brand URL. Chains Firecrawl → GPT Image 2 → Seedance 2.0 via the Higgsfield MCP. Brand-bound to DESIGN.md tokens; outputs ready to upload to LinkedIn Campaign Manager and Meta Ads Manager.
Doctrine inherited (Step 7 — 0626 rollout)
Output complies with:
output-tenets.md— the seven tenetsoutput-simplicity.md— length caps, three-layer source placement, robot-tells bandesign-production.md— DESIGN.md contract + banned visual patterns (composes with The Iron Law below)- Step 6 calibration: see [[feedback_execution_doctrine_refinements_step6]]
Refinements applied to this skill:
| Code | Refinement | How it lands in ad-creative |
|---|---|---|
| R1 | Source placement (three layers) | Ads are end-customer-facing. No sources block in the ad copy or on the creative. Run manifest captures prompts + sources for QA; not surfaced in the deliverable. |
| R2 | Single-doc-with-toggles | Multi-format ad pack ships as one run-folder + one summary doc with per-platform toggles (LinkedIn 1.91:1 / 1:1 / Meta 1:1 / 4:5 / 9:16). Not 8 separate sub-docs. |
| R3 | Product-update tone | Ad copy frames as "[Product] now does X" not "we're thrilled to announce." Even hero-launch ads default to even-keeled product framing. |
| R5 | Blog as voice anchor | When the campaign has an anchor blog, the blog's opening line becomes the hero headline across LinkedIn + Meta + variants. Verbatim. Cross-platform voice consistency is what makes the pack read as one campaign. |
| R6 | CTA hierarchy | Market-facing ads → sign-up or trial primary. Blog/whitepaper as fallback for cold awareness. Never both. Retargeting / warm-base ad sets → product-action CTA. |
| R9 | Action-oriented section names | Brief sections + asset captions verb-led. "How to start with [Product]" beats "Get started." |
Triggers
Run this skill when:
- A campaign needs creative for both LinkedIn and Meta simultaneously and a human design queue is too slow.
- A new launch needs a same-day asset pack to A/B test angles before committing to bigger spend.
- A client wants UGC-style video without booking a creator shoot.
- The operator already has a brand-kit DESIGN.md and wants on-brand AI generation, not a generic stock-photo look.
Do NOT run when:
- The output is going to a human designer for production — use
/ad-creative-briefinstead. - There's no DESIGN.md and no brand URL — the skill needs at least one to bind brand identity.
- The campaign needs a single hero shot with art-directed humans (still a creator-shoot job; AI UGC is not photoreal humans).
- The Higgsfield MCP isn't installed (see Prerequisites).
The Iron Law — voice-locked
ASSETS ARE BRAND ARTIFACTS, NOT AI OUTPUTS.
Every asset the skill produces must read as if it were made for this brand specifically — not as a generic AI render. The DESIGN.md token contract is the floor, not the ceiling.
No exceptions:
- Off-palette colors → reject the render, re-prompt with explicit hex tokens.
- Fabricated stats, testimonials, or customer logos in overlays → strip them. Mark
[NOT AVAILABLE]if the brief implies a number we can't source. - Generic UGC personas with no ICP grounding → re-roll with the brand-kit's audience description.
- Em dashes in ClientCo overlays (or any client with a no-em-dash rule) → re-prompt without them.
- Mixing rounded + sharp corners in the same asset → reject; pick one per
design-production.mdDo's/Don'ts.
Prerequisites — Higgsfield MCP
This skill depends on the Higgsfield MCP for GPT Image 2 + Seedance 2.0 calls (Higgsfield also exposes Soul V2, Veo 3.1, Kling 3.0, Flux 2 from the same server), plus the already-installed Firecrawl MCP for the brand-brief stage.
Higgsfield uses an HTTP MCP server with OAuth-based authentication — no API keys to manage. Auth happens once per machine, on first tool use, via a browser sign-in to your Higgsfield account.
One-time setup (run before first invocation):
# 1. Install the Higgsfield MCP at user scope (persists across all worktrees):
claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcp
# 2. Verify it registered:
claude mcp list | grep higgsfield
# higgsfield: https://mcp.higgsfield.ai/mcp (HTTP) - ! Needs authentication
# 3. Authenticate via OAuth:
# Open a Claude Code session and run the /mcp slash command.
# Higgsfield will print an auth URL — open it in a browser, sign in to your
# Higgsfield account, copy the returned code, and paste it back in Claude.
# Alternatively, authentication is triggered automatically the first time
# any mcp__higgsfield__* tool is called.
# 4. Confirm tools are surfaced:
# In a fresh Claude Code session, look for tools named
# mcp__higgsfield__* in the deferred-tool list.
Startup check (the skill runs this on every invocation):
The skill calls a no-op Higgsfield tool (e.g., mcp__higgsfield__list_models or equivalent) before stage 1. If the call returns "tool not found" or "needs authentication," the skill aborts with:
"Higgsfield MCP not detected or not authenticated. Run
claude mcp add --transport http --scope user higgsfield https://mcp.higgsfield.ai/mcpthen/mcpto authenticate. See SKILL.md."
No partial runs — if the MCP is missing or unauthenticated, no Firecrawl credits get spent either.
Cost gate: Higgsfield credit spend is non-trivial. The skill estimates total run cost before stage 2 (image generation) and asks the user to confirm if estimated cost > $5 USD-equivalent. See the premium reference."
Inputs
| Input | Description | Required |
|---|---|---|
| Brand URL | The brand's website root (e.g., https://ClientCo.com) — Firecrawl scrapes for hero copy + visual hints |
Required |
--client {slug} |
Routes output to projects/consulting/active/{slug}/paid/execution/. Triggers brand-kit lookup at brand/{MMYY}-brand-kit.md |
Recommended |
| DESIGN.md (brand-kit) | Token frontmatter (colors, typography, components) — the brand contract | Recommended; falls back to transient brief if absent |
--brief {path} |
Path to a pre-existing /ad-creative-brief output to consume — overrides Firecrawl extraction for angles + copy |
Optional |
--copy {path} |
Path to a /linkedin-ads-copy or /google-ads-copy output to use as text overlays + UGC scripts |
Optional |
--mode |
full (default), statics-only, ugc-only, animate-only |
Optional |
--platforms |
Comma-separated subset of linkedin,meta (default both) |
Optional |
Process
Seven-stage deterministic chain (Stage 7 is an optional pre-ship gate). Each stage has explicit MCP tool, model pick, prompt template, expected output, and failure mode. Full reference in the premium reference.
| # | Stage | MCP / Model | Output |
|---|---|---|---|
| 1 | Brand brief | Firecrawl scrape + DESIGN.md merge | brief.md |
| 2 | Hero static | mcp__higgsfield__generate_image → marketing_studio_image (default; ads-tagged Higgsfield model) or gpt_image_2 (text-heavy variants) |
per-aspect static PNGs |
| 3 | Text overlay | Sharp/Jimp local OR flux_kontext for context-aware edit |
static-with-overlay PNGs |
| 4 | Animated hero | mcp__higgsfield__generate_video → seedance_2_0 (default; identity-consistent) |
6s MP4 per aspect ratio |
| 5 | UGC persona | mcp__higgsfield__generate_image → soul_2 (UGC + portrait specialty; supports soul_id for identity persistence) |
persona portrait PNG |
| 6 | UGC clips | mcp__higgsfield__generate_video → seedance_2_0 or wan2_7 (lip-sync) |
9:16 + 1:1 MP4s |
| 7 | Virality predict | mcp__higgsfield__virality_predictor (optional gate) |
scores written to manifest; ranks A/B variants |
Mode flags skip stages: --statics-only runs 1→3; --ugc-only runs 1+5+6; --animate-only runs 1+2+4. --no-virality skips Stage 7.
Aspect-ratio honesty: LinkedIn 1.91:1 renders as native 16:9 (no model supports 1.91:1 directly; LinkedIn auto-crops 16:9 cleanly). Meta 4:5 static is native to marketing_studio_image but NOT gpt_image_2. Meta 4:5 video is not native to any model; spec ships 1:1 + 9:16 for video, not 4:5.
DESIGN.md token contract — voice-locked
Every prompt to GPT Image 2 + Seedance 2.0 cites brand tokens explicitly. Same contract as the rest of the marketing surface.
| Required in every prompt | Forbidden |
|---|---|
Token-cited colors (e.g., colors.primary: #1A1C1E) |
Free-form color names ("a warm blue") |
Token-cited typography (typography.headline-lg.fontFamily: Inter) |
Off-brand font names from the model's defaults |
Brand-kit's signature_elements (logo treatments, photographic style) |
Generic stock-photo aesthetics |
| Voice overrides (e.g., ClientCo no-em-dash) on every overlay | Em dashes / banned tokens leaking through |
Authority: Full integration contract in .claude/rules/design-production.md (auto-loaded). If DESIGN.md doesn't exist, the skill builds a transient brief from Firecrawl and flags the gap — it does NOT invent token values.
Design cycle (post-authoring phases)
After producing the happy-path output, walk these phases before ship. Each references the shared design-quality library at ../../../meta/catalog/design-reviewer/the premium reference. Run /design-reviewer` as the final ship-ready gate.
- Layout —
layout-tenets.md(rhythm, alignment, density across aspect ratios) - Distill —
distill-principles.md(strip generative noise; one message per asset) - Typeset —
typeset-principles.md(overlay measure, leading, scale — esp. on 9:16 mobile) - Polish —
polish-principles.md(16 details; verify caption legibility on muted video) - Cognitive load —
cognitive-load-tenets.md(when overlay copy is dense) - Delight —
delight-patterns.md(1–3 motion moments per video; nothing more) - Final review — run
/design-reviewer(5 dimensions × 0–4, P0–P3 severity)
Skip Harden + Onboarding — they apply to code/app output, not ad creative.
Anti-Hallucination Guardrails
- Never invent customer stats. "2,000 advisers trust us" needs a verified brand-kit source. Otherwise strip from overlays.
- Never invent token values. If DESIGN.md doesn't define a needed color, pause — do not guess.
- Never approximate platform specs. Use the table in
process.md. - Never let the model improvise voice. Pass brand-kit voice overrides into every prompt (ClientCo: no em dashes; ClientCo brand-name = single word).
- Never strip captions from video. 85% of LinkedIn + Meta video is watched muted; auto-generate caption tracks.
- Never ship without /design-reviewer pass. Review-gate 3 is the floor for external publication.
Quality
Self-evaluation checklist (token coverage, aspect-ratio compliance, overlay character limits, caption presence, voice override compliance), worked examples (ClientCo + ClientCo), failure-mode triage, and the rerun-loop pattern in the premium reference.
Integration with Other Skills
Upstream (consumes)
| Skill | What it provides | Required? |
|---|---|---|
brand-kit |
DESIGN.md tokens | Recommended (falls back to transient brief) |
ad-creative-brief |
Angles + visual concepts | Optional (overrides Firecrawl extraction) |
product-messaging |
Value props, taglines | Optional (used in overlays) |
linkedin-ads-copy / google-ads-copy |
Headlines for overlays | Optional |
Downstream
Terminal deliverable — assets upload directly to LinkedIn Campaign Manager / Meta Ads Manager. Run through /design-reviewer as the final gate. Track results via /paid-audit after spend begins.
Final ship gate
Run /premortem --output before ship. See /premortem skill for the 5 execution domains (will-it-resonate / will-it-convert / will-it-stay-on-brand / will-stakeholder-push-back / will-it-degrade-over-time) and output template.
Trivial-case escape: ## Premortem\nNo failure modes — trivial change satisfies the contract for genuinely trivial outputs.
Persuasion & stickiness pass
Output complies with persuasion-and-stickiness.md — Cialdini's 7 persuasion levers + Heath's SUCCESs. Deploy the 1-2 Cialdini levers that fit the reader's barrier (never all seven; every lever must be TRUE), run the SUCCESs diagnostic (Simple / Unexpected / Concrete / Credible / Emotional / Stories) over the near-final draft, then the rule's pre-ship gate.