Fichiers de la Skill (11)
---
name: aident-ecommerce-product-images
description: Plan, generate, and adapt product visuals for domestic and international ecommerce and social platforms through Aident Loadout-managed Fal and language-model capabilities. Cover listing main images, details, Amazon A+, feeds, and vertical content when the user wants multiple platform versions from real product photos without separate image-provider API keys.
---
# Aident Multiplatform Product Images
## Loadout capabilities
- Upload a local product or logo reference to temporary Aident storage with <action-tag>direct:file_storage_tools:file_create_upload_url</action-tag>, then obtain a fresh provider-readable URL with <action-tag>direct:file_storage_tools:file_get_download_url</action-tag> immediately before use.
- Inspect the current image-understanding catalog with <action-tag>direct:fal:fal_list_media_analysis_models</action-tag>, then extract visible product evidence with <action-tag>direct:fal:fal_analyze_media</action-tag>.
- Organize observed evidence and user-confirmed facts into a structured fact table and image plan with <action-tag>direct:llm_inference:llm_generate_structured</action-tag>. Supply text evidence; this Action does not itself inspect images.
- Inspect current image-edit models, fields, and pricing with <action-tag>direct:fal:fal_list_image_edit_models</action-tag>. Generate reference-based listing images, or run a supported background-removal endpoint, with <action-tag>direct:fal:fal_edit_image</action-tag>.
- When current destination rules must be checked, search and read official platform results with <action-tag>composio:firecrawl_tools:firecrawl_search</action-tag>. Inspect its live schema, Vault readiness, and cost first. This reference is optional when the user supplies current authoritative requirements or only requests a draft without a compliance claim. If official rules cannot be verified, mark the affected delivery as provisional rather than ready to upload.
Read [Aident Actions and file handling](references/aident-actions.md) before selecting a model, uploading local files, or constructing Action inputs. These tags identify separate executable Actions; reading this Skill never authorizes their execution or a paid batch.
Read the [platform-format index](references/platform-formats.md) for any named platform, placement, cross-platform adaptation, or size export. Load only its linked profiles relevant to the current task. Canvas suggestions are design starting points, not current upload requirements.
## Prerequisites
1. Read the installed `aident-skill` and its `references/loadout.md` in full. Follow the setup entrypoint above if Aident is unavailable.
2. Start from live CLI help, capability discovery, exact Action schemas, and Vault status. Do not rely on an old model name, price, or input field copied from this package.
3. Require at least one real product photo or an accessible product-image URL. Confirm that the user is authorized to use the supplied material. Request usable logo artwork if a logo is required; never reconstruct or invent it.
4. Check current platform, placement, country, and category requirements through the referenced Aident web-research Action, another live-discovered suitable Aident Action, or a current authoritative specification supplied by the user. A request for compliance, ready-to-upload assets, or latest formats, and any advertising, Amazon A+, or store-decoration module, requires current official-rule verification. Record the exact source, check date, and scope. Do not claim compliance from memory or assume main-image rules also apply to secondary images or A+ modules. If requirements cannot be verified, state the unresolved constraint and prepare a provisional proposal. A sign-in-only seller requirement needs an authorized user-provided specification or screenshot, not guessed rules.
5. Before promising deterministic resizing, composition, or export, check whether the host has a local renderer and an image-inspection tool. Inspect their supported formats and output path. If missing, discover an appropriate Aident Action and quote any additional cost, or report the exact export limitation. Do not call remote Sandbox execution caller-local or promise that it is free.
## Workflow
### 1. Collect references and required facts
Once product imagery is available, ask once for the following missing information. Do not repeat anything clearly provided in the current conversation:
1. Product name or category.
2. Sales or publishing platforms, allowing one or more. If the intended surface is ambiguous, clarify its placement: product page, store-decoration module, feed, Story, Reels, short-video cover, or advertisement.
3. Target country or region.
4. Image language.
5. Confirmed specifications such as length, width, height, weight, material, color, and model. Collect only what the user can confirm.
6. Exact mandatory image text, and whether a logo must appear. If required, obtain usable logo artwork.
Unknown specifications may be marked `unknown`. They do not block planning, but they must never appear in image copy or product claims. Missing platform, target market, or image language requires a follow-up; never silently default to Amazon, the United States, or English. If multiple markets or languages are requested, map them explicitly to placements rather than silently applying one combination to all exports.
Do not require the user to choose image count, image types, selling points, audience, use cases, visual style, composition, marketing copy, or per-platform export dimensions. Infer a proposed direction from the references, confirmed facts, current placement requirements, and target market, then present it for review.
### 2. Understand the product and confirm the fact table
Upload local references to temporary Aident storage using the linked file-handling contract. List current media-analysis models and select one suitable for product description or visual question answering, rather than an unrelated classifier. Inspect its schema and preflight the exact input before calling the media-understanding Action. If analysis or structured planning has a nonzero Aident cost, disclose it and obtain approval for that bounded preliminary work before execution; approval for image generation does not silently include an unquoted analysis call.
Use media understanding to extract visible facts. Then use structured LLM output to organize:
- Product category, visible color, shape, and material appearance.
- Packaging, accessories, and visible text.
- Specifications and selling points explicitly confirmed by the user.
- Plausible use scenarios, clearly labeled as scenarios rather than product facts.
- Unsupported claims that must be omitted.
Keep four evidence classes separate: `user-confirmed facts`, `visually observed facts`, `inferred use scenarios`, and `unproven performance claims`. A material's appearance does not establish its composition, and a scene does not prove dimensions or performance. Model analysis is provisional evidence; inspect the source imagery and ask the user to correct material uncertainties.
Derive selling points, audience, and scenario copy only within the evidence boundary. Omit unsupported load capacity, anti-slip performance, durability, rotation angle, compatibility, certification, efficacy, and comparative advantages. Do not invent or solicit unsupported claims merely to fill a layout.
Show a compact fact table, proposed core selling points, and the unknown claims to omit. Obtain the user's confirmation or corrections before paid image generation.
### 3. Route platforms and build the delivery matrix
Split each selected platform into actual placements. Treat product main images, detail pages, feed images, vertical covers, store modules, and ads as distinct destinations even within one platform. Use the applicable profile linked from the platform-format index. For an unlisted platform, build a task-specific profile with the same fields rather than refusing the task or guessing another platform's rules.
Create one row per output or placement variant with platform, country or region, image language, purpose, aspect ratio, pixel dimensions, format, file-size limit, text policy, safe area, background policy, official-rule evidence, and verification status. Distinguish platform requirements from proposed design choices and unresolved account-specific requirements. Never treat a suggested canvas as evidence that an upload will be accepted.
Limit coverage to the user's selected platforms and placements. Present the matrix together with the master plan and cost estimate in one review after fact confirmation; do not add a separate matrix-approval loop. Official-rule research is separate from enabling a generation model's web search and never introduces new product facts.
### 4. Plan masters and image sets
Choose a justified image count and mix based on the platform, product complexity, reference quality, and confirmed facts. Suitable components include a compliant main image, key selling-point images, lifestyle scenes, detail close-ups, usage steps, dimensions or specifications, package contents, verified model or color comparisons, and a brand closing image or CTA. Include only useful components; avoid repetitive filler. Use only confirmed numbers in dimension images, and do not imply that scene props are included accessories.
For multiple placements, plan a small reusable master set:
- An isolated-product master: high resolution, complete silhouette, and no baked-in marketing text. Reuse it for approved clean-background main images and specification layouts.
- A scene master: keep the product inside the planned safe region with enough surrounding context for distinct square, portrait, vertical, or landscape compositions.
- An infographic background: keep the product and decorative background separable where possible. Lay out text, supplied logos, dimension arrows, and CTAs deterministically with local HTML/CSS, Canvas, or SVG when supported.
Preserve product proportions. For small aspect changes, use a composition-safe crop, padding, or background extension. For a large change or missing scene context, quote generative outpainting or regeneration only when needed and supported by the selected live model. Do not blindly crop a finished square layout into a vertical image. Product masks, layers, or editable sources are deliverables only when actually created and inspected, not inferred from a flattened generation.
Propose the visual style, composition, hierarchy, selling points, audience, scenarios, and short copy. Respect the chosen platform and category's limits on overlays, logos, props, claims, backgrounds, and module dimensions. Keep promotional elements in placements that allow them. Without usable logo artwork, use an unbranded design and invent no brand or certification.
Use the explicitly chosen image language. Adapt units, reading direction, and wording to the target market; unit conversions must derive from confirmed measurements without inventing precision.
For every proposed master and derivative, show its purpose, placement, dimensions or aspect ratio, composition, exact copy, reference roles, and generation prompt or deterministic adaptation method. Include the delivery matrix and the itemized cost estimate from the next step. Obtain one confirmation of that complete plan and the quoted batch expense. The user only needs to correct inaccurate facts or unwanted choices; do not ask them to write the design brief, dimensions, or selling points themselves.
### 5. Choose models and quote costs
Before each generation configuration, list current Fal image-edit models and inspect their current capabilities, input fields, and pricing. Follow pagination or narrow the search when the first page lacks a suitable model.
- For subject preservation and ordinary scenes, prefer the current default edit endpoint or an available Seedream precision-edit endpoint when its live capabilities suit the task.
- For dense text or demanding English typography, prefer a current endpoint with typography support and reference-image inputs.
- For white or transparent backgrounds, prefer a dedicated background-removal endpoint exposed through the edit catalog. Verify real alpha for transparency; composite a white background locally when the destination requires white.
- Avoid embedding long passages in a generation prompt. Generate a clean reference-based background and use local HTML/CSS, Canvas, or SVG for deterministic typography when the host supports it. Render to the final image format and inspect the result; an HTML layout alone is not a deliverable image.
Call `capabilities preflight` with the exact Action and input for every distinct configuration, including image count, dimensions, resolution, and model-specific options. Present the user-facing USD estimate, whether it is exact or estimated, and the bounded quantity for the entire batch. Use the Aident quote as authoritative rather than multiplying an old provider price. Include the representative image in the total and identify any separately approved preliminary costs.
Itemize official-rule research and product analysis or LLM planning, paid master generation, paid outpainting or repainting, and deterministic adaptation/export. The last category may add no model-generation fee, but hosted execution, rendering, storage, or another Action may still be billable. Quote those charges when applicable; never label an entire remote workflow free. Reuse the same quote for identical configurations only while it remains valid; re-preflight changed inputs and check the approved total before continuing.
Execute only after the user explicitly agrees to that batch expense. Pass any runtime acknowledgement or credit-confirmation fields only for the exact approved call. A new model, larger batch, or changed quote requires approval for the difference. Failures and retries may incur additional expense; explain and obtain agreement to the incremental cost before a paid retry.
### 6. Generate and adapt platforms
- Keep the same approved product references and factual constraints in every generation. State explicitly that silhouette, color, logo, components, and packaging must remain consistent.
- Request only the images needed in the current Action call. Use model-specific `num_images` or equivalent fields from the live schema; do not assume one model's fields apply to another.
- Generate one representative image and inspect it before continuing the approved batch, unless the user explicitly requests immediate batch generation. If it fails the approved direction or product-identity checks, stop to propose a correction and quote any paid replacement.
- Keep model web search disabled unless the user explicitly authorizes it and permits outside facts to enter the creative process. Consult platform documentation separately; platform research does not authorize product-model web search or new product claims.
- Reuse approved masters without stretching the product. For clean or solid-background images, prefer canvas padding and proportional resizing. Use generative expansion only when missing scene content requires it, after quoting the selected model configuration.
- Reflow text, logo placement, and visual hierarchy for every target ratio. Keep the product, mandatory text, logo, and any confirmed price clear of the verified placement's UI overlays and safe-area exclusions. If a placement has no verified safe-area specification, mark the layout provisional and request the actual preview or account specification before claiming compliance.
- Export marketplace main images separately from social marketing assets. Do not carry social CTAs, promotion badges, or decorative copy into a placement that requires a clean product image. Never invent a price, offer, discount, or urgency claim.
- Give every derivative a master ID and adaptation record. Record crop or padding, output dimensions, text version, placement, and any additional paid call so that identical content and costs are not counted twice.
- Check both the Action envelope and provider result for errors. A queue receipt or request ID is pending work, not a finished image. Reconcile asynchronous jobs to terminal success before reporting delivery.
- After a timeout or missing response, inspect Aident audit history and any returned provider-status mechanism first. Recover completed result files rather than dispatching a duplicate paid request. Never fabricate a polling Action or bypass Aident with raw provider credentials.
- If a required capability becomes unavailable, retain the approved fact table and plan, report the precise blocker, and stop the dependent step.
### 7. Verify and deliver
Download all completed outputs promptly into a task-specific directory organized by platform and placement. Keep approved masters separately and use clear names such as `01-main`, `02-key-feature`, and `03-lifestyle` with the correct extension. Inspect every image against the approved source, matrix, and plan for:
- Product identity, silhouette, colors, logo, components, packaging, and included accessories.
- Exact mandatory copy, spelling, image language, readability, and the absence of unknown specifications or unsupported claims.
- Actual pixel dimensions, aspect ratio, crop, background, real alpha when required, file format, color mode, byte size, placement safe areas, and UI-obscuration risk. Read dimensions, mode, and bytes with an available inspection tool rather than relying only on visual appearance. If the final bytes cannot be inspected, mark that verification incomplete.
- Current destination rules, their region/category/placement scope, and remaining seller-account checks. Recheck changes between research and final export when the destination or rules have changed.
- Consistency of the entire set, with no duplicate or unnecessary images.
Regenerate only failed images, after approval of any additional cost. Deliver the platform directory, a full-set preview or contact sheet, a manifest linking every output to its master and matrix row, the verified official-rule URLs and check dates, remaining seller-account checks, and an actual Aident cost summary derived from execution receipts or audit records. Distinguish paid preliminary work, generation, expansion/repainting, execution/export, and approved retries. If actual settled cost is unavailable, say so instead of reporting the preflight estimate as actual expense. Use the host's supported media renderer and verify that the files open.
Use the manifest fields and selected-platform folder examples in the platform-format index. Do not create empty deliverable folders for unrequested platforms. Retain original masters and export derivatives separately so a later platform change does not require regenerating a correct product image.
## Boundaries
- This Skill replaces the cloud execution dependency on Meitu and `product-shots`; it does not claim to be their official model or service.
- Managed credentials avoid separate provider-key setup. Fal and other paid Actions still consume Aident credits.
- When the user asks only for a plan, prompts, or preflight, do not run paid generation. Complete that requested scope without requiring generated outputs.
- Do not process unauthorized private material or expose signed upload/download URLs, tokens, or raw credentials in public copy. Temporary URLs are execution inputs, not permanent delivery links.
- Deliver assets for review; uploading a listing or publishing it to a store requires a separate user instruction.
- Multiplatform coverage means the selected platforms and placements, not every known platform. Profile canvases are design suggestions and never permanent upload rules or guarantees of platform approval.
