SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsmedia-higgsfield-product

media-higgsfield-product

Design & creation

View repository

Updated 10-15 21:03

About this skill

Higgsfield Product Shoot (media-higgsfield-product)

moai-media | Brand/product visual 10-mode determination and generation (core: media-higgsfield-core)

Overview

"Make me a product photo" is not a single kind of request. Studio cuts, Pinterest pins, and ad packs all differ in composition, ratio, number of images, and prompt structure. This skill first determines the request as one of 10 modes, then assembles it using that mode's prompt structure.

The call contract and cost preflight follow the core:

  • Call contract: ../media-higgsfield-core/references/call-schema.md
  • Job/cost/readback: ../media-higgsfield-core/references/job-lifecycle.md

⚠️ Known limitations (read first)

The official Higgsfield CLI has a product-photoshoot dedicated backend prompt enhancer. The server holds the photo terminology and structure templates for each mode and assembles the prompt for you.

This enhancer is not exposed to the MCP connector. Even when querying the marketplace app list, the app is not there (observed at time of writing). Therefore this skill cannot call the enhancer, and assembles the prompt structure for each mode directly and passes it to a general-purpose image model.

As a result:

Item This skill (MCP) Official CLI
Mode determination ✅ Same ✅
Interview flow ✅ Same ✅
Prompt assembly Skill does it directly Server enhancer
Fidelity Cannot guarantee Official optimization
Additional installation None CLI installation/login required

The official documentation states that bypassing the enhancer and calling the image model directly "makes the results noticeably worse." If maximum fidelity is the goal guide the user to the official CLI skill (higgsfield-product-photoshoot) path. This skill's advantage is that it can be used right now with no additional installation, not superior quality.

Do not hide this limitation from the user.

Trigger keywords

product photo, product cut, studio cut, cutout background, lifestyle cut, staged cut, Pinterest pin, hero banner, top banner, carousel, slide ad, ad creative, Meta ad, virtual fitting, wearing cut, lookbook, conceptual, floating product, splash, restyle, seasonal version

10 modes

Mode When
product_shot Product on a neutral/studio/catalog background
lifestyle_scene Real environment, hands, motion, mood
closeup_product_with_person Tight crop including part of the hand and face — beauty demo, usage scene
moodboard_pin Vertical 2:3 Pinterest moodboard
hero_banner Website, email, campaign horizontal header
social_carousel IG, LinkedIn, Facebook connected 3–10 images
ad_creative_pack Meta, TikTok, Pinterest, Google static ad variant bundle
virtual_model_tryout AI model wearing/using
conceptual_product Surreal, CGI, floating, splash, sculptural
restyle Changes the aesthetic, mood, and season of an existing image

Decision rules and tie-breaking are references/modes.md. The pre-interview types (A–F) are references/interview.md.

Workflow

Step 1 — Mode determination

Not by surface keywords but by intent. If both apply, the more specific one wins. For details, see references/modes.md.

Step 2 — Interview (up to 4 questions)

Classify the user's situation into types A–F and ask only the questions for that type. Do not exceed 4 questions in one message. Skip questions whose answers are already given in the context. Ask with choices attached, not open-ended questions.

This skill does not ask the user directly — it returns the slot as a blocker and the orchestrator asks.

Step 3 — Prompt assembly

Assemble using the per-mode structure (references/modes.md's per-mode components). Since there is no enhancer, this step determines quality.

Step 4 — Live lookup and cost preflight

models_explore(action:'get')Check the target model's constraints (aspect ratio, resolution) with get_cost: trueand check credits with

Step 5 — Approval gate

Stop before credits are spent — follow the core §Paid Generation Approval Gate as is. Since this skill has no enhancer and assembles the prompt directly,(Step 2 caution), if the assembled result diverges from the user's intent, that cost is wasted as is. The gate is the place that prevents that.

  • [HARD] The approval screen shows the full assembled prompt. The "do not return the full prompt" in §Precautions below is a rule about result reporting, not a rule about the approval screen. Before money is spent, show what the money is being spent on.
  • Multiple variants (social_carousel·ad_creative_pack included) show number of images × unit price = maximum total creditstogether and get approval once per batch.

Step 6 — Generation and delivery

Call only with approved values. Reference images are passed as media_id or job_id(URLs rejected). Results are delivered briefly as a list of image URLs.

Multiple variants

When requesting a number of images, vary the preset, lighting, angle, and palette from each other — do not make it a reconstruction of the same prompt.

social_carousel과 ad_creative_packmeans the number of images is the number of slides/variants. These two modes Fix the visual system of the previous slideTherefore, even when introducing variations, the palette, typography, and compositional grammar are shared.

Output format

## Higgsfield 제품 비주얼 결과
- 판정 모드: [mode] — 근거: [걸린 신호]
- 사용 모델: [models_explore로 확인한 실제 id]
- 비율·장수: [aspect] · [count]
- 비용: [get_cost가 반환한 credits]
- 결과:
  - [URL]
  - [URL]
- 서버 조정(adjustments): [있으면 그대로 보고]

In the results report, do not return the full assembled prompt — what is wanted is the output. However, the Step 5 approval screen is an exception, and there the full text is shown(since credits have not yet been spent).

Precautions

  • State the limitations first. If the task requires high fidelity, guide them to the official CLI path.
  • Do not exceed the limit of 4 interview questions.
  • If the mode is chosen incorrectly, everything else goes wrong — the determination is the first quality gate.
  • Do not invent and use names outside the 10 modes.
  • Creating from text alone without product photos significantly reduces fidelity. Recommend uploading photos first.
  • Do not composite competitors' trademarks or others' copyrighted works.
  • Staging that implies performance or efficacy different from the actual product may constitute a violation of the Act on Labeling and Advertising — in a commerce context, moai-seller:commerce-ad-claim-compliance-krhand it over to .

Related skills

Skill Timing
moai-media:media-higgsfield-core Core: invocation contract, cost, namespace
moai-media:media-higgsfield-image Alternative: general-purpose image not requiring mode determination
moai-media:media-higgsfield-identity Prerequisite: model and mascot consistency reference
moai-seller:commerce-detail-page-image Follow-up: detail page layout
moai-seller:commerce-product-photo-brief Prerequisite: shooting brief
moai-seller:commerce-ad-claim-compliance-kr Follow-up: legality of advertising expressions
moai-designer:design-brand-visual Adjacent: brand system consistency

Source

  • Higgsfield Skills (official agent documentation) — higgsfield-product-photoshoot v0.12.0 (MIT). Basis for the 10-mode definitions, determination rules, tie-breaking, and interview types.
  • Live MCP marketplace app list lookup — basis for the backend enhancer not being exposed. Evidence tier: primary (negative observation).
  • The prompt assembly structure is reverse-engineered from the official mode definitions and does not guarantee the same output as the official enhancer.