Social Marketing Asset Generator

This skill's differentiation angle: baked-in typography via structured JSON captioning, not the generate-text-free-then-composite-in-HTML/CSS pattern every confirmed competitor in this space uses. Ideogram renders the scene and the headline/subhead/CTA typography together in one call, leaning into its specific strength — reliable in-image text rendering with full control over placement, size, weight, and color. See references/typography-baking-discipline.md for the full writeup of why this beats the composite-afterward approach, and what it costs (garbled-text risk, caught and regenerated against rather than accepted).

This is a prompt-composition skill — same family as logo-prompting, brand-identity-sheet, and moodboard-generator — not an MCP-orchestration skill. Its center of gravity is composing a legible, on-brand, on-copy JSON caption; the MCP calls it chains (generate_image → optional generate_images_bulk fan-out → optional edit_image touch-up) exist to serve that composition, not the other way around.

It is not a copywriting or ad-strategy skill. It accepts literal copy — headline, subhead, CTA, offer/price if any — as intake and renders it. It never runs AIDA/PAS strategy frameworks, writes ad copy unprompted, or invents a claim, price, or offer that wasn't supplied. A rendered image with a fabricated discount figure or an unapproved claim baked into the pixels is a real failure this skill guards against, not a nice-to-have.

Before you start

• references/composition-spec-format.md — read before composing any JSON caption. It names the six ad-creative element roles and gives a full worked example.
• references/anti-slop-discipline.md — read before running the pre-generation gate. It lists the universal and ad-specific bans the gate scores the caption against.
• references/platform-dimension-map.md — read before resolving any placement's aspect_ratio. It maps platform-native pixel specs to Ideogram's supported enum and classifies the fit.
• references/typography-baking-discipline.md — read before the legibility review or any touch-up request. It has the legibility rules, the one-message-hierarchy rule, and the edit_image-vs-regenerate decision table.

Workflow

1. Intake

Gather brand truth: the project's existing brand.md if one exists; otherwise the brand name, colors, and a product/offer image if relevant. Gather the literal copy strings — headline, subhead, CTA, and offer/price if any — and the target placement(s).

If literal copy isn't supplied and the ask is vague ("make an ad for our sale"), ask for the exact strings before drafting anything. Never invent a claim, price, or offer detail — this is the one input worth stopping for, since a fabricated number baked into pixels is a real failure, not a stylistic shortcut.

2. Resolve platform dimensions

For each requested placement, look up its row in references/platform-dimension-map.md.

• Exact fit → proceed silently.
• Near-fit → proceed, but mention the approximation in the final save-what-you-made summary (Step 8).
• Mismatch (e.g., the LinkedIn Personal/Company Banner row) → proceed like Near-fit, but disclose it to the user up front, before generating — not just in the Step 8 summary — since the nearest enum only shares the native ratio's general orientation and does not reach its magnitude, a larger gap than a Near-fit approximation.
• Severe mismatch (currently only the Leaderboard row) or any placement not in the table with no close ratio match → tell the user up front which placement(s) will be an approximation — or, for a severe mismatch, recommend handling that placement outside this skill — before generating anything.

3. Compose the JSON caption

Build the caption per references/composition-spec-format.md's six element roles: background/scene, product/subject, headline, subhead/tagline, CTA, brand-mark. Place each text element's bbox inside the resolved platform's safe zone from platform-dimension-map.md. Apply the legibility rules from references/typography-baking-discipline.md while drafting, not after: max words per line, text size relative to canvas, contrast-checked color pairing, and the one-message-hierarchy rule (headline dominant, subhead secondary, CTA tertiary).

4. Anti-slop gate

Score the drafted caption against references/anti-slop-discipline.md's four-axis pre-generation gate — cliché avoidance, message hierarchy, copy fidelity, safe-zone placement — before generating. Any axis scoring under 3 means revise the caption now, not after generating. Only send to generate_image once all four axes score 3 or higher.

5. Generate

Single placement: call mcp__ideogram__generate_image with the composed prompt (the JSON caption, passed as the prompt string), the resolved aspect_ratio (and resolution only where platform-dimension-map.md lists an exact pixel match for that row), style_type: "DESIGN", rendering_speed: "QUALITY".

Multiple placements from one concept: call mcp__ideogram__generate_images_bulk with one prompt per resolved platform variant — same underlying concept, text-element bboxs adjusted per platform's safe zones. If bulk generation errors or is unavailable (it's a Pro-only feature), fall back to sequential generate_image calls per platform and tell the user that's what happened and why.

6. Legibility review

Inspect the generated image for garbled or illegible rendered text — Ideogram's known failure mode on dense in-image copy, named in references/anti-slop-discipline.md's ad-specific bans. If the text is garbled, regenerate from the JSON caption with adjusted text-element sizing/spacing (larger bbox, fewer words per line, per typography-baking-discipline.md's legibility rules) rather than accepting a broken render.

7. Touch-up loop (optional)

For a localized fix, follow the edit_image-vs-regenerate decision table in references/typography-baking-discipline.md:

• mcp__ideogram__edit_image for a genuinely localized change — recolor a CTA button, swap a logo mark, fix one word.
• A fresh generate_image call from a corrected JSON caption for anything bigger — a full headline, layout, or placement change.

State explicitly which route is being taken and why, so the user isn't surprised by a freshly regenerated asset when they asked for a quick touch-up.

8. Save

Write the final JSON caption(s), prompt(s), the platform/dimension mapping actually used (including any disclosed near-fit approximations or severe-mismatch recommendations from Step 2), and the resulting image URL(s)/response_id(s) to the project's existing output location. Check for an existing logo-explorations/ or branding/ folder first and match whatever convention brand-identity-sheet, logo-prompting, and moodboard-generator already use for that project — same pattern moodboard-generator/SKILL.md follows. Per the toolkit's "No Context Lost" habit, nothing generated in this skill should live only in the conversation.

Error handling

• No literal copy supplied and the ask is vague — ask for exact headline/CTA/offer strings before drafting; never fabricate a claim, price, or offer detail.
• Requested placement has no close aspect_ratio match — tell the user explicitly which placement(s) will be a near-fit or mismatch approximation before generating (for a Mismatch, call out that the gap is larger than a Near-fit approximation since the nearest enum only shares the native ratio's general orientation, not its magnitude; or, for a severe mismatch, recommend handling that placement outside this skill). Don't silently ship a wrong-ratio asset as if it matched spec.
• Garbled or illegible rendered text — treat as a failed render, not an acceptable result. Regenerate with adjusted text-element bbox/sizing rather than delivering it.
• generate_images_bulk unavailable (non-Pro account or a tool error) — fall back to sequential generate_image calls per platform and tell the user that's what happened, rather than silently failing the whole request.
• edit_image requested for a change bigger than a localized touch-up — don't force it through edit_image. Explain that a full copy/layout change goes through a fresh generate_image call from a corrected caption instead, per its own strict single-image-edit contract.
• Existing brand logo must appear pixel-exact and unaltered — flag that Ideogram generation can't guarantee pixel-fidelity reproduction of an existing logo asset; recommend the user composite that exact logo file in afterward with their own tooling, rather than claiming to bake in a logo it can only approximate.

Reference files

• references/composition-spec-format.md — the ad-creative element vocabulary (six element roles) and a full worked example.
• references/anti-slop-discipline.md — universal and ad-specific bans, plus the four-axis pre-generation gate.
• references/platform-dimension-map.md — the placement table (Instagram, Facebook, LinkedIn, X, YouTube, Pinterest, generic display) and the Fit classification (Exact/Near-fit/Mismatch/Severe mismatch).
• references/typography-baking-discipline.md — legibility rules, the one-message-hierarchy rule, and the edit_image-vs-regenerate decision table.
• examples/ — worked ad-creative jobs go here as they happen; empty for now.