creative
AI creative production engine for product photos, product video, social graphics, talking heads, ad creative, and video ads cut from real or generated footage in code and rendered to MP4 for free, through one tiered pipeline with a vision-QA loop that looks at every asset before you see it. Ad creative is built from 336 live Meta ads we viewed (21 formats with real examples, a hook bank, a script that pulls competitors' live ads from the Meta Ad Library, a batch plan that mixes long-running stills with fresh UGC, and a visual check every ad must pass). Use whenever visual assets or ad creative need to be made, edited, planned, or batch-produced for a brand or campaign.
How do I install this agent skill?
npx skills add https://github.com/boringmarketer/boring-marketing-skills --skill creativeIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The 'creative' skill is a professional-grade production engine for generating marketing assets. It integrates securely with major AI providers and uses local scripts for research and model verification. No significant security risks were detected beyond the inherent risk of indirect prompt injection from competitor ad data.
- Socketpass
No alerts
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Creative Engine
One engine for every visual asset a brand ships. The judgment: match motion and style to positioning, spend pennies finding direction before spending dollars on finals, and never present an asset you haven't looked at yourself. Done = downloaded files, QA-checked, brand-consistent, named so downstream skills can find them. This is the floor, not the ceiling. When the situation argues for a better path, take it and say why.
Context
Only if ./brand/, ./brands/ or MARKETING.md exists (nothing to load otherwise), run:
python3 ${CLAUDE_SKILL_DIR}/../start-here/scripts/brand_context.py --for creative
(If the script or python3 is unavailable, read the files named in reads: directly,
section-anchored where specified. If none exist, proceed standalone. Ask only for
what the task strictly requires.)
If learnings.md holds creative priors (hook formats, style winners), apply them. Mention one only when it changed the work, as a single line in the note.
Setup: 30 seconds, once
- Token: check
$REPLICATE_API_TOKEN(env or.env). Missing → prompt-only mode, defined inreferences/api-workflow.md§6: same workflow, deliverable is the exact prompt + model + settings instead of the file. Never block on the token. - Stack freshness: if
./brand/creative-stack.mdis missing or older than the registry'sttl-days, runbash ${CLAUDE_SKILL_DIR}/scripts/verify_models.shbefore the first paid call. It rewrites./brand/creative-stack.md(a different file from thestack.mdthat start-here keeps) with a pass/fail table. - Brand kit: read
./brand/creative-kit.md. Missing → for a one-off asset, proceed and offer to build it after; for campaign work, build it first (below).
What we use
None of this is required. With no key, prompt-only mode writes the exact prompt, model
and settings for any tool you already have. When you want the files made for you, this is
the stack we run, one pick per job. Keys are listed in TOOLS.md; payloads and per-model
prices live only in references/MODEL_REGISTRY.md.
- Images. We use GPT Image 2.6 (OpenAI): Flare for fast drafts, Sunburst for finals and
for keeping a product consistent across edits. It holds the
image-complex-promptandimage-instruct-editroles (exact layout, edits to a real photo, native-looking phone photos, animated GIFs). It is pay per image, through the OpenAI API (OPENAI_API_KEY) or Replicate. Everyday stills stay onimage-default. Ad text is set in HTML over the image, not generated inside it. Without a key, the prompt-only deliverable works in any image tool. - Video. We use Seedance 2.5 (ByteDance) through Higgsfield or Replicate, pay per
clip. That means 2.5 where your provider lists it, otherwise 2.0. The registry's
verified rows are Seedance 2.0, and 2.5 sits there as a "newer, verify before use" row
until it passes the registry's checks. For ads, Higgsfield Marketing Studio is our pick
for presenter or UGC-style video and for exploring many static formats fast (an AI
presenter presents, never reviews); Seedance 2.5 covers b-roll, foley and talking-object
skits. Setup, test results and what its API covers are in
references/ad-formats.md. Without a key, you get the prompt and settings for any video tool. - Voiceover and talking heads. We use ElevenLabs for the voice, for cloning and a
consistent voice across a campaign. It has a free tier, then plans (
ELEVENLABS_API_KEY). Without it, the lip-sync model's built-in voice or any quick text-to-speech tool covers a script test. - Video ads cut in code. We use HyperFrames (by HeyGen, Apache-2.0): the agent
writes HTML and a GSAP timeline, and the CLI renders MP4 on your machine. Free, no key,
no per-render fee; it needs Node 22+ and FFmpeg. It cuts real or generated footage
together with captions, review cards, offers and end cards, and renders hook variants
from one project.
Remotion (React to MP4) is the alternative; check its license for your company size.
Mode:
modes/motion-ads.md. - One key for many models. Replicate is the path the registry's payloads and the
async pattern are written for (
REPLICATE_API_TOKEN). Higgsfield is the one-key alternative: one API for 50+ image and video models, Seedance included, pay as you go (HF_API_KEY_IDandHF_API_KEY_SECRET). On Higgsfield keep the role and the prompt, and take the parameters and calls from its own docs.
References: when to load what
| File | Load when |
|---|---|
references/model-routing.md | Before model selection: job→model defaults, the i2v/i2i/t2v decision, style capsules, moodboarding, platform map, when to leave Replicate. |
references/MODEL_REGISTRY.md | Before ANY generation: the only source of payloads and prices. Resolve roles here. |
references/api-workflow.md | First generation of a session: async pattern, download-within-1h, file naming, campaign manifest, QA mechanics, handoffs. |
references/VISUAL_INTELLIGENCE.md | Constructing prompts for anything that matters: visual psychology, anti-slop checklist (Part 4 powers the QA rubric), style taxonomy, consistency techniques. |
references/ad-formats.md | Choosing, making or iterating ad creative: 21 formats from 336 live Meta ads with real examples and counts, the three routes (real, model, code), Higgsfield, production lessons, the translation for any business type, GPT Image 2.6 GIFs. |
references/hooks.md | Writing the first second of an ad: 12 opening patterns seen in live ads, when each fits, fill-ins for any business, making hook variants. |
references/motion-ads-starter.html | Starting a video ad cut in code: a checked 9:16 composition (hook clip with a text box at frame 0, body with captions, review card over footage, end card on a real photo, hook variants) that passes the render checks and the Reels, TikTok and Shorts safe zones. |
scripts/ad_library.py | Looking at what competitors run before planning ads: pulls live US Meta Ad Library ads for a brand or category, downloads them, writes an index and a board (ad-creative mode, step 1). |
Modes
Route by what's being made. Each mode file carries the craft for that asset type; they all share the registry, the API workflow, and the QA loop.
| Request smells like | Mode file |
|---|---|
| Product stills, e-commerce shots, editing a real product photo | modes/product-photo.md |
| Reveals, orbits, demos, any product motion | modes/product-video.md |
| Posts, thumbnails, quote cards, content calendars | modes/social-graphics.md |
| Presenters, spokesperson, UGC-style presenter videos (never posing as a customer), lip-sync | modes/talking-head.md |
| Paid ads, UGC ads, ad concepts, static ads, hooks, ad matrix cells, A/B variants, which ad formats to make | modes/ad-creative.md |
| Video ads cut from real or generated footage (UGC cut, hands demo, review card over footage, before/after, reply to a comment, screen recording), captions on a video, hook variants | modes/motion-ads.md |
| Anything else visual | No mode file: brand kit + registry + this file |
Quick Mode: a specific single-asset request ("16:9 hero shot of the bottle on walnut") skips mode ceremony. Construct the prompt, generate, QA, deliver, one-line what's-next.
Routing by ROLE: never by model ID
Load references/model-routing.md before model selection. It carries the job→model
judgment and the input-modality (i2v/i2i/t2v) decision. Payloads and prices appear
ONLY in references/MODEL_REGISTRY.md; this file and the modes speak in roles, and
the registry's Role Map resolves them at call time. When models change, the registry
changes and nothing else does.
STILL IMAGE
├─ default → image-default (iterate with seeds where supported)
├─ shipping at 4K → image-premium-4k
├─ editing a real photo / multi-reference fusion → image-default with image_input
├─ complex composition / dense prompt / exact in-scene text → image-complex-prompt
├─ typography-dominant flat graphic (>6 words) → image-text-render
├─ logo / icon / needs real vector → image-vector
└─ user names a model → use it (look up its payload in the registry)
VIDEO: the tier pipeline (below) is the default path
├─ drafts → video-draft
├─ UGC / ad-native (handheld, imperfect-real) → video-ugc
├─ polished finals → video-production
├─ character/reference consistency → video-reference
├─ locked-off product shot → video-locked-camera
└─ explicit "hero" → video-premium (3-model parallel)
SPEECH
├─ lip-sync default → lipsync-default
├─ cheap script tests → lipsync-budget
├─ video-production output → lipsync-kling
└─ animate a still photo into a speaker → digital-human
CODE-MADE VIDEO ADS (no model, no key)
└─ cut real or generated footage with captions, review cards, end cards
→ modes/motion-ads.md (text-only motion only for SaaS screen recordings
or an offer bumper, never the main ad for a physical product)
Don't ask the user which model to use. Pick the role. Model names, seeds and per-asset cost go in the manifest; the chat shows one line of total spend in the note.
The tier pipeline (default for all video)
Cheap drafts find the direction; production money only follows a chosen direction.
1. SOURCE FRAME Best video starts from an approved still (I2V beats T2V: you
control frame one). Generate/approve the image first when practical.
2. DRAFT 3-5 motion variants on the draft tier (draft-tier cost per MODEL_REGISTRY, 720p,
~30s). Different camera moves / energies, not five near-clones.
QA each, present pre-ranked.
3. PICK User picks a direction (or asks for another draft round).
4. PRODUCTION Regenerate the winner(s) on the production role at full quality.
Same prompt skeleton, same seed discipline where supported.
5. PREMIUM ONLY on an explicit "hero" request: run the premium trio in
parallel on the winning prompt, QA all three, present pre-ranked.
Never silently escalate to premium.
Images skip the tiers: generate on image-default, iterate by changing ONE variable
per round, reusing the seed where the model supports it so you can attribute the change.
Cost discipline: under $1 total: just generate (no confirmation ritual). $1-2: state the cost in one line and proceed. ≥$2: confirm before running. Always record actual spend in the campaign manifest.
Vision QA loop: runs after EVERY generation
You look at everything before the user does. No raw URL dumps.
- Download immediately (URLs die in 1 hour; api-workflow.md §1).
- Look:
Readthe image; for video extract first/mid/last frames via ffmpeg (mechanics in api-workflow.md §2; thumbnail fallback if ffmpeg is absent). Motion ads: read the renderer's snapshot contact sheet before rendering, then frames from the final MP4. - Score 1-10, silently, against a rubric assembled from: the brand kit's palette/mood/ avoid-list, the anti-slop checklist (VISUAL_INTELLIGENCE.md Part 4), and the request's own spec (ratio, text space, focal point, platform).
- Below 7: name the failure, fix THAT in the prompt, regenerate. About 2 retries is the practical ceiling; past that the prompt frame or model family is the constraint, so present the best attempt with its flaw named instead of burning budget.
- Present the asset that passes: what it is and where it fits, in a line. Name a residual flaw (a label warp at 3s) in one line only if one remains. Multi-model comparisons arrive pre-ranked with a reason per option, no scores or model names.
An asset that stays below 7 is shown with its flaw named, never passed off as clean. Scores
go in the manifest (qa_score) so /performance-review can later correlate them with
results; the client does not see them.
Brand kit: ./brand/creative-kit.md
The visual DNA every mode reads before generating: colors (hex + names), typography direction, photography/illustration preference, overall mood, the avoid-list, logo path, and any saved artifacts (style locks, social templates, approved presenters).
Building it: ask for a site/social URL first and propose a kit from what you observe;
fall back to direct questions only for what you can't infer. Writes/updates follow the
brand-memory spec (${CLAUDE_SKILL_DIR}/../start-here/references/brand-memory.md: auto-version old file, 3-line change summary, update brand.yaml).
Using it: palette and mood terms are injected into every prompt; avoid-list items are standing guardrails; saved style locks and presenter specs are reused verbatim rather than re-explored.
Style exploration: optional, not a toll booth
For a NEW project with no locked style and real campaign scope, offer (don't impose) a direction exploration, default five: the category-safe one, its opposite, one borrowed from another industry, one emotion-first, one convention-breaker; fewer when the brief is narrow. Generate small, present with strategy/vibe/risk per direction, let the user pick or combine ("lighting from 2, composition from 4"), then write a style lock (palette, exact lighting phrase, composition rules, technical reference, anti-patterns) into the campaign directory and reference it in every subsequent prompt.
Skip exploration when: a style lock or template already exists, the request is specific, it's a follow-up asset, or the user says skip. Quick Mode always skips.
Consistency at scale = brand prefix + locked phrases + fixed seeds + reference images (techniques: VISUAL_INTELLIGENCE.md Part 10).
Batch & campaign scale
3+ assets → write the campaign manifest BEFORE generating
(api-workflow.md §4: per-asset role, model, prompt, seed, cost, qa_score, status).
Generate in parallel waves of 5-8 (rate limits), updating the manifest per asset. A
failed batch resumes from the manifest. Skip done, retry failed once, run
pending. Wall-clock for a wave = its slowest member.
Output
Files-first, per the api-workflow.md conventions: assets under creative-output/…
(or the campaign's own tree), versioned names, ad assets named by matrix cell ID,
one line per asset appended to ./brand/assets.md. Model, prompt, seed, cost and QA
score per asset live in the manifest, not in chat.
Delivery summary follows the standard contract:
## <DELIVERABLE>+ one line of date and what the work is for- The assets, finished pieces first, each with its path, what it is, where it fits, and a one-line flaw only if one remains
- One short Sources and assumptions note (10 lines max): total spend in one line, what the work rested on (the kit, or a site read to build one), what to confirm, and what was not made (prompt-only mode, no ffmpeg so thumbnails only)
**Files saved:**exact paths with (new)/(updated)**What's next:**(below)
Ad sets: the ads plus a short test plan. If a creative brief came from another skill,
hand the creative_delivery block (api-workflow.md §5) to that skill; never show it in
the client-facing text.
When the user edits your work
Apply the edit, then make it stick: save a gold case to ./brand/gold/<this skill>/ and
one to three yes/no checks to ./brand/checks.md, and confirm in one line. Before handing
over any work, answer the checks the context block loaded, silently, and fix every no.
How: ${CLAUDE_SKILL_DIR}/../start-here/references/taste-loop.md.
What's next
/content-atomizer: cut the approved assets into platform-native posts (~10 min)/paid-ads: if these were matrix cells, build the campaigns around them (~15 min)/direct-response-copy: landing page or email copy to pair with the visuals (~15 min)/performance-review: after the assets have run, map results back to cells and QA scores so the next round is smarter
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/boringmarketer/boring-marketing-skills/creative">View creative on skillZs</a>