Higgsfield Content Factory

Higgsfield Content Factory — Campaign Orchestration Pipeline

A 5-stage pipeline that runs a full ad campaign on top of Higgsfield Marketing Studio: Research → Plan → Generate → Publish → Report. Where higgsfield-marketing-studio is the engine reference (what the presets, parameters, hooks, and avatars are), Content Factory is the orchestration layer that drives a whole campaign turn by turn — researching trends, planning a dated content calendar, generating in batches, scheduling to ads, and reporting cost savings.

Translated from Adil Aliyev's higgsfield-content-factory source skill, per the v3.7.13 / v3.7.16 translation precedent. All Marketing Studio API ground-truth — preset slugs, parameter schema, hook/setting/avatar handling — lives in ../higgsfield-marketing-studio/SKILL.md and is NOT restated here. This sub-skill carries the pipeline scaffolding the marketing-studio sub-skill deliberately deferred. The publish + cost-report tail (Stages 4–5) lives in the companion publish-and-report-workflow.md satellite.

1. Two operating principles

Content is UGC-first. Every campaign defaults to a heavy share of UGC-family content — talking-head clips, honest reviews, ASMR close-ups, unboxing reveals. Hyper Motion, TV Spot, and Wild Card stay available but off the default mix. Every video idea must be producible inside one live Marketing Studio preset within its duration cap (see marketing-studio § 3 / § 7 for the canonical preset list and limits).

The interaction is button-driven. Every clarifying question is asked as a 2–4 option button choice (one round-trip), never a free-form "type your answer" for navigation, confirmation, or routing. Free-form typing is reserved for content the user must originate — and even then, always offer a smart default they can accept with one click. The only non-button steps are attaching a product image or pasting a product URL (both click-based file actions).

User-facing language rule (HARD). The user is not a developer. Do not narrate technical mechanics — no MCP tool names, no UUID resolution, no "running parallel searches," no enhanced-prompt internals. All of that runs silently. Send one plain-language stage banner when each stage starts, a brief "Stage N done — [deliverable]" between stages, then the next banner. The user sees a clean narrative arc, not a tool log. If the user asks "how does it work under the hood," then explain.

Stage banners

Send one banner per stage, in plain language (no tool names):

2. Onboarding — one message, all buttons, no pauses

Ask the onboarding questions in a single button round-trip at the start. Do not ask them sequentially; do not pause between onboarding and the product request.

• A — Higgsfield connection: "Yes — connected" · "Not yet — I'll connect now" · "Skip — research only."
• B — Starting stage: Stage 1 full pipeline (needs a product) · Stage 2 build content plan (have a brief) · Stage 3 generate now (have a plan) · Stage 4 schedule (content ready). A product image with no other context defaults to Stage 1.
• C — Video volume: 50 · 100 (recommended) · 150 · 200 · Other (any number). Store as [VIDEO_COUNT].
• D — The product: "Attach your product image OR drop a URL — that's all I need to start." Skip D if a product is already attached. For Stage 3/4, swap D for "drop your existing content plan."

Send A + B + C as buttons and the product-attach prompt in the same message; once answered, proceed straight to the chosen stage with no extra confirmation.

Compute the per-format split silently the instant [VIDEO_COUNT] locks in — do not announce the breakdown here. It surfaces later, woven into the Stage 1 brief as a consequence of what the research showed, so it reads as a trend-driven choice rather than a config card.

3. The campaign mix — 5 UGC formats (even split)

Every campaign distributes evenly across five viral UGC formats. Cinematic secondary types (Hyper Motion / TV Spot / Wild Card) stay off-default — used only when the user explicitly asks.

Allocation math. per_format = floor(VIDEO_COUNT / 5). If VIDEO_COUNT doesn't divide evenly, distribute the remainder one-per-format from format 1. Examples: 100 → 20/20/20/20/20; 50 → 10×5; 17 → 4/4/3/3/3.

This is the default mix and is user-overridable. The default share is a craft choice (UGC-first), not a platform rule.

Format → preset routing

Each format maps to a Marketing Studio preset. Route preset selection through the canonical mechanism documented in ../higgsfield-marketing-studio/SKILL.md § 3 (slug naming is canonical there; routing mode lives on a separate MCP call). At a glance:

• Formats 1 (Entertainment), 2 (Street Interview), 5 (ASMR) → the UGC preset; what differentiates them is the system hook, the setting, the audio flag, and the prompt content — not a separate preset.
• Format 3 (Unboxing) → the Unboxing preset.
• Format 4 (Product Review) → the Product Review preset.

Do not hardcode hook/setting/avatar UUIDs or invent a generate_video.mode slug. Enumerate hooks/settings live (marketing-studio § 4 live-enumeration discipline) and pass the resolved IDs at generation time.

Concept seeds (vary within each format)

No two videos in the same format should be the same concept. Pull from seeds like: blind taste try, "$100 to try it" street challenge, will-it-pour (Entertainment); "rate it out of 10," "sing for the product," blind opinion → brand reveal (Street Interview); trio reveal, ribbon-pull solo drop, subscription-box drop (Unboxing); two-ingredient test, fridge-ranking, 7-day diary (Product Review); macro cap-unscrew + glug pour, condensation slide, spoon-clink ice-drop (ASMR).

ASMR is a style, not a preset

Render ASMR as the UGC preset with audio on, an intimate low-noise setting (Kitchen / Bathroom / Bedroom), usually no system hook, and a prompt focused on close-up handling sounds (cap unscrew, pour, glass clink, condensation). No talking; caption-only or silent open.

4. Stage 1 — Trend research & viral ideas

All research from live web searches; every idea validated against live Marketing Studio capability; UGC-first weighting.

1. Probe capability silently. Pull the current preset / hook / setting picklists live (marketing-studio § 4 / § 5 enumeration). Cache the hook + setting IDs for Stage 3. Never narrate this.
2. Auto-detect product & niche from the image or URL — do not ask the user to confirm.
3. Run trend research — a spread of web searches across the product's niche on the major platforms, plus at least two source-page fetches, run silently.
4. Synthesize a Viral Content Brief (UGC-first): 15+ idea cards, each tied to a format + preset, with the campaign's format breakdown woven in as a consequence of what the research surfaced.
5. Producibility self-check before adding any idea — every idea must be renderable inside one preset within the duration cap. Ideas needing lip-sync from non-human characters, multi-character coordinated dialogue,

   15s single clips, or split-screen are out (or routed to an escape- hatch model — see § 6).

6. Approve via buttons before moving to Stage 2.

5. Stage 2 — Video content plan

Turn the approved brief into a dated content plan: one row per video ([VIDEO_COUNT] rows), bucketed by the 5 formats, each with its format, preset, concept, hook/setting intent, and a calendar date. Render it as a polished, saved document deliverable the user can review and approve before any generation fires. Interleave dates so formats rotate across the calendar rather than clustering.

6. Stage 3 — Generate in Marketing Studio

The point of failure-avoidance here is the per-batch gate: never fire a whole campaign at once.

1. Upload the product silently and register it (marketing-studio § 6).
2. Resolve hooks / settings / avatars at runtime per the cached picklists; never hardcode IDs.
3. Generate one preset batch at a time, and ask before each batch fires with a button gate ("Generate the 20 UGC Entertainment videos now?"). This keeps the user in control and prevents a runaway spend.
4. No on-screen text in prompts — Marketing Studio renders text unreliably; keep copy out of the generation and add it in post or via the image pack.
5. Image asset pack (optional) — generate social posts / hero banners / product stills via GPT Image 2.0, scaling the count from [VIDEO_COUNT] (see ../higgsfield-gpt-image-2/SKILL.md for prompt craft; the count split skews toward social posts, then hero, then with/without-people stills). A standalone image-pack mode skips the videos entirely.
6. Handle failures by re-resolving the picklist and retrying the single failed batch, not the whole campaign.

Escape hatch — ideas Marketing Studio can't render

If an idea genuinely needs lip-sync, longer narrative, or multi-shot continuity, label it "Outside Marketing Studio" and route to Wan 2.7 (synchronized audio + character-consistent), Veo 3.1 (cinematic, audio- friendly), Cinema Studio Video 3.0 (cinema-grade), Seedance 2.0 (reference- driven, multi-SKU), or Kling 3.0 (multi-shot, motion transfer). Use sparingly — UGC-first means most ideas stay inside Marketing Studio.

7. Stages 4–5 — Publish & report

Stage 4 (schedule & publish to Meta Ads) and Stage 5 (cost-comparison report) live in the companion publish-and-report-workflow.md satellite. That tail carries the most connector-dependent and estimate-based material, so its caveats are isolated there.

8. Hard rules (consolidated)

• Button-driven navigation; free-form typing only for user-originated content, always with a one-click default.
• 5-format even split as the default mix; user-overridable.
• Producibility check before any idea ships to the plan.
• Per-batch generation gate — confirm before each batch fires.
• Hooks are visual scene templates, not verbal copy; never confuse a hook with caption text.
• Between-stage confirmation before continuing.
• Defer all API ground-truth to marketing-studio; never restate or invent preset slugs, parameter shapes, or UUIDs here.
• Keep internals silent unless the user asks how it works.

9. Source acknowledgment

Translated from Adil Aliyev's higgsfield-content-factory source skill (997 lines). The 5-stage pipeline, the UGC-first 5-format campaign taxonomy with concept seeds, the even-split allocation math, the button-driven UX discipline, and the per-batch generation gate are the net-new orchestration material the v3.7.13 marketing-studio translation deliberately deferred. Marketing Studio API specifics are not re-documented here — they live in ../higgsfield-marketing-studio/SKILL.md, which also reconciled several source-corpus API claims (preset slug naming, avatar handling, the absence of a generate_video.mode slug); this sub-skill honors those reconciliations and routes all API decisions through it. Default format shares and concept seeds are craft opinion, presented as overridable defaults, not platform rules.