ads-suite-pipeline
ads-suite-pipeline
Required inputs
- One of these starting points:
- a source pack from
$ads-info-gatherer - an existing approved anchor design or approved partial suite the user wants to extend
- a source pack from
- Target vehicle models and requested sizes
- Approved retail truth set for copy/disclaimer content
- Any approved reference designs, prior approved suites, or campaign-specific style guardrails
- Budget/approval expectations if the run may require many paid image-generation retries
Workflow
- Create a run folder:
- Working:
logs/car-ads-designer/ads-suite-pipeline/<task-id>/ - Final:
artifacts/car-ads-designer/ads-suite-pipeline/<task-id>/ - Start a
README.mdimmediately with the request summary, chosen task id, known source materials, and current status
- Working:
- Confirm the campaign truth set is ready.
- If the user did not provide a usable source pack or approved anchor asset, run
$ads-info-gathererfirst - Copy or link the approved source inputs into
source-pack/inside the run folder - Load the active workflow, brand playbook, or campaign brief when client-specific rules apply
- If the user did not provide a usable source pack or approved anchor asset, run
- Decide the starting mode.
fresh-suite: no approved anchor exists yet, so choose one anchor model and start at1200x1200continuation: an approved anchor design already exists, so reuse it and continue with the remaining sizes/models without re-opening already approved work unless the user asks
- Set the anchor plan.
- The anchor size is
1200x1200unless the user already supplied an approved anchor asset for a different starting point - Do not start any other size before the anchor design for the current model is approved
- Do not start later model suites before one full reference suite is approved
- The anchor size is
- Generate or reuse the clean anchor image.
- For
fresh-suite, generate the clean base image with$fal-nano-banana-2-image-gen - For
continuation, reuse the approved clean/base asset and document where it came from - Generate images without letters
- No dealer text, no pricing, no disclaimer text, no logo text blocks
- Keep enough clean space for later copy and legal overlay
- Preserve the intended vehicle identity, trim, and viewpoint as closely as the source pack requires
- Pause for review before broad creative changes or unusually large paid rerun batches
- For
- Review the clean image and loop until pass.
- Load
references/review-checklist.md - Check vehicle fidelity, composition quality, clean zones, seasonal/regional fit, and suite consistency
- Write
review.jsonandreview.md - If the result is
fail, write a narrow regeneration brief and generate again - Do not continue until the current clean/base asset passes
- Load
- After the clean/base asset passes, create final copy layers.
- Build copy as SVG, not raster text
- Use exact approved copy only
- Never invent, beautify, shorten, or paraphrase VINs, prices, trims, or legal text unless the source pack explicitly approves that variation
- Save copy sources, SVG exports, and layout notes under
models/<model>/<size>/copy/
- Overlay the SVG copy onto the approved base image.
- Export one composite candidate for the current size
- Review the final composite against the same checklist, plus text fidelity, logo safe space, disclaimer readability, and footer collisions
- If failed, adjust layout/overlay and review again until pass
- Keep failed candidates in the working folder and only place approved finals in the deliverables area
- Finish the rest of the sizes for the same model one by one.
- Use the approved anchor design as the reference for the rest of that model's size suite
- For each size: clean image -> review loop -> SVG copy -> overlay -> final review loop
- After one full model suite is approved, use it as the style anchor for the remaining vehicle models.
- Other model suites may then run simultaneously if useful
- Keep the suite visually consistent with the approved anchor suite
- Write a complete handoff.
- Record the source pack or approved anchor input, anchor model, anchor size, requested sizes, pass/fail history, and final deliverables
- Make it obvious which files are drafts, which are blocked, and which are approved to deliver
References
- Load
references/review-checklist.mdwhenever reviewing either the clean image or the final composite. - Use the active workflow, approved source pack, or campaign brief for client-specific sizes, approval rules, and branding constraints.
Outputs
- Working folder:
logs/car-ads-designer/ads-suite-pipeline/<task-id>/README.md— live status, request summary, next step, and blocked itemssource-pack/— copied or linked source materials used in the runmodels/<model>/<size>/clean/— prompts, generations, regeneration briefs, and clean-image reviewsmodels/<model>/<size>/copy/— SVG copy assets, approved text source, and layout metadatamodels/<model>/<size>/final/— candidate composites, final reviews, and rejected variants
- Final deliverables:
artifacts/car-ads-designer/ads-suite-pipeline/<task-id>/README.md— handoff summary, anchor suite, approved outputs, and any explicit blockerssource-pack/— final linked or copied truth-set bundle used for the approved runmodels/<model>/<size>/final/— only approved deliverables safe to hand off
Keep intermediate generations, failed candidates, and review loops in logs/. Keep only approved handoff materials in artifacts/.
Defaults & rules
- Default to
1200x1200as the first anchor size unless the user already supplied an approved anchor design to continue from. - Default to finishing one model suite before starting the next, even if multiple models were requested.
- If copy/facts are incomplete or contradictory, stop and resolve the truth set before generating more finals.
- If exact copy does not fit, adjust layout instead of altering the facts.
- Treat generated text inside an image as unusable until it has been rebuilt from the approved truth set as SVG.
Definition of done
- The anchor model has an approved full size suite
- Every requested size for every requested model has either an approved final asset or an explicit blocked status
- Review files explain every failed attempt and every approved deliverable
- The working folder and final handoff each include a
README.md - The final handoff makes it obvious which files are safe to deliver
artifacts/contains only approved deliverables and their supporting handoff notes
Safety / quality checklist
- Do not skip the
1200x1200anchor approval - Do not generate final ad text inside the base image stage
- Do not proceed to the next size before the current size passes
- Do not start later model suites before one full reference suite is approved
- Treat VIN, trim, pricing, and disclaimer mistakes as blockers
- Keep failed or uncertain outputs out of the final deliverables folder
- Do not approve layout changes that compromise disclaimer readability or logo safe space
More from lingkaix/smartworkers
smart-skill-maker
Create or improve SmartWorkers-style skills from a workspace-local `skills/` source tree, then apply them to Codex with `npx skills`. Use when you want one guided tool for new skill creation, skill upgrades, or SmartWorkers skill-maintenance work while keeping repo conventions aligned.
22ads-info-gatherer
Use when the user needs a dealer car-ad source pack built from live web sources or a provided worker-pack folder, with VIN-anchored retail truth, source assets, and provenance preserved for downstream generation and review.
7fal-nano-banana-2-image-gen
Generate images from text prompts using fal.ai Nano Banana 2 (fal-ai/nano-banana-2). Use when you need fast text-to-image variants for ads, concepts, or backgrounds.
6workspace-setup
Initialize a SmartWorkers-style agent workspace with repo-root guidance, `logs/`/`temp/`/`artifacts/`, a local `skills/` source tree, ignore rules, config templates, and the required global `mise` plus `npx skills`, `skill-creator`, and `smart-skill-maker` bootstrap completed in one turn by default. Use when starting a new agent workspace, bootstrapping a fresh project folder for repeatable agent work, or standardizing README, WORKFLOW, AGENTS, and skill-management flow before adding more automation.
5fal-veo3-image-to-video
Generate videos from a reference image using fal.ai Veo 3.1 image-to-video (`fal-ai/veo3.1/image-to-video`). Use when you need short video variants (ad b-roll, product shots, storyboards) from a still image, with final outputs saved to `artifacts/` and request/response logs saved to `logs/`.
1fal-qwen-image-edit
Generate and edit images from reference images using fal.ai Qwen (`fal-ai/qwen-image-edit-2511`). Use when you need image-to-image edits/variations, background swaps, style alignment, or multiple aspect-ratio outputs (e.g., square + 16:9) from the same prompt.
1