Scenarios

Starter scenarios across the Higgsfield skills. Each is one user request + what the agent should do + how to score it.

These exist to be run by a human (or by another agent acting as the user) in a fresh session with skills installed.

Scenario 1 — Cheap photorealistic image (generate)

User request:

Make me a quick photorealistic image of a fox in a snowy forest.

Expected behavior:

• Picks nano_banana_2, flux_2, or another photorealistic-default model.
• Does NOT pick a Soul model (no face mentioned).
• Does NOT pick gpt_image_2 (overkill for "quick").
• Submits via higgsfield generate create <model> --prompt "..." --wait (one-shot create+poll, no separate wait step).
• Stays silent until done (no "checking status..." spam).
• Delivers ONE URL with a short summary.

Score:

• Pass: correct model class, single URL out, no internal narration.
• Partial: correct model but excessive narration ("calling the API now…").
• Fail: wrong model class, multiple jobs submitted, broken URL.

Scenario 2 — Image-to-video animation (generate)

User request:

Animate this photo into a 5-second clip with the camera slowly pulling back. [attached: still.jpg]

Expected behavior:

• Picks kling3_0 (default for image-to-video) or seedance_2_0.
• Uses --start-image still.jpg.
• Uses motion verbs in the prompt ("pulls back", "ambient motion") not redescribed scene.
• --duration 5.
• Delivers ONE video URL.

Score:

• Pass: correct model, --start-image used, prompt focuses on motion.
• Partial: redescribed scene in prompt instead of motion-only.
• Fail: text-only video model picked (ignored the photo), wrong duration, no --start-image.

Scenario 3 — Soul training (soul)

User request:

Train my Soul on these 8 photos. Call it "founder". [attached: 8 photos]

Expected behavior:

• Picks --soul-2 (default).
• Submits higgsfield soul-id create --name founder --soul-2 --image ... --image ....
• Polls silently. Soul training takes 15–45 min.
• Delivers: "Soul founder ready. Use in generate with --soul-id <id>."
• Does NOT print the raw reference_id in chat (UX rule).

Score:

• Pass: correct variant, all 8 photos passed, terminal success message clean.
• Partial: prints raw reference_id (against UX rule).
• Fail: wrong variant, fewer photos passed than provided, polling spam.

Scenario 4 — Soul + Generate chain

User request:

Use my Soul "founder" (already trained) to make a cinematic portrait of me on a Tokyo street at night.

Expected behavior:

• Looks up the existing Soul reference (asks user for the id, OR detects from a workspace state file if one exists).
• Picks text2image_soul_v2 or soul_cinematic (the cinematic variant fits the "cinematic" word).
• Passes --soul-id <id>.
• Does NOT re-train Soul.
• Delivers ONE URL.

Score:

• Pass: correct chaining, no re-training, cinematic variant chosen.
• Partial: re-trained Soul unnecessarily.
• Fail: ignored Soul, used a generic photo-real model.

Scenario 5 — Pinterest pin (product-photoshoot)

User request:

Make a Pinterest pin of my candle [attached: candle.jpg]. Cottagecore mood.

Expected behavior:

• Picks --mode moodboard_pin.
• NOT lifestyle_scene or product_shot.
• Calls higgsfield product-photoshoot create --mode moodboard_pin --image candle.jpg --prompt "...".
• Does NOT call higgsfield generate create gpt_image_2 ... directly (must go through the prompt enhancer).
• Asks ≤4 short questions (count, mood, anything to emphasize). Mode is obvious from the request.
• Delivers ONE URL or a short bulleted list if --count > 1.

Score:

• Pass: correct mode, routed through product-photoshoot (not direct gpt_image_2).
• Partial: asks too many interview questions.
• Fail: wrong mode, bypassed the prompt enhancer.

Scenario 6 — Hero banner with use case (product-photoshoot)

User request:

Hero banner for my landing page showing my serum being applied.

Expected behavior:

• Picks --mode hero_banner (banner format wins over closeup-with-person tie-breaker per SKILL.md).
• Default aspect ratio appropriate for hero banner (16:9 or wider).
• Skips Type-A interview because mode is obvious.

Score:

• Pass: hero_banner mode, correct aspect ratio.
• Partial: closeup_product_with_person (specific genre tie-break that loses to banner format here).
• Fail: random other mode.

Scenario 7 — Marketing Studio UGC video from URL (generate)

User request:

Make a 15-second UGC ad for https://shop.example.com/sneakers, 9:16.

Expected behavior:

• Calls higgsfield marketing-studio products fetch --url ... --wait first.
• Then lists/picks an avatar (higgsfield marketing-studio avatars list).
• Then higgsfield generate create marketing_studio_video with --mode ugc, --duration 15, --aspect_ratio 9:16, --wait.
• Asks one question at a time. Does NOT batch-ask (avatar + product + mode + duration upfront).

Score:

• Pass: full 3-step Marketing Studio flow, single-question phases.
• Partial: batch-asked questions upfront.
• Fail: tried to use generic kling3_0 instead of marketing_studio_video.

Scenario 8 — Language detection

User request: (in Russian)

Сгенерируй мне фотореалистичную картинку лисы в зимнем лесу.

Expected behavior:

• Detects user_language = ru.
• Replies in Russian for status, questions, summary.
• Keeps technical flags English (--model nano_banana_2, --aspect_ratio 16:9).
• Picks nano_banana_2 or flux_2.

Score:

• Pass: Russian for human-facing text, English for flags.
• Partial: half-translated (e.g. translates "16:9" to "16 на 9").
• Fail: replies in English when user wrote Russian.

Scenario 9 — Vague request (product-photoshoot Type F)

User request:

Make me something cool for my brand.

Expected behavior:

• Routes to higgsfield-product-photoshoot Type F interview.
• Asks 2–3 disambiguating questions (what product, goal, reference image?).
• Does NOT submit a generic image immediately.

Score:

• Pass: structured 2–3 labeled questions.
• Partial: too many open-ended questions ("what's your brand about?").
• Fail: submits a generic image without asking.

Scenario 10 — Don't invent model names

User request:

Use Higgsfield's new "supernova_v9" model to make a sci-fi cityscape.

Expected behavior:

• Notices the model name is unfamiliar.
• Runs higgsfield model list to verify.
• Reports back: "supernova_v9 isn't in the catalog. Suggested alternatives: …".
• Does NOT submit with a fabricated model name.

Score:

• Pass: verified before submitting, correct fallback.
• Partial: submitted anyway and surfaced the API error.
• Fail: hallucinated some other model name without verification.

Scenario 11 — Virality Predictor video scoring

User request:

Analyze this ad video and tell me whether the hook is strong enough to hold attention. [attached: ad.mp4]

Expected behavior:

• Picks Virality Predictor (brain_activity).
• Uses --video ad.mp4.
• Does NOT ask for a prompt; the video is the input.
• Treats the result as a text score report, not a generated video/image.
• Delivers overall score, peak hook second, sustain score, region highlights, and a business interpretation of attention/virality potential.
• Links the Open report URL for visual inspection.

Score:

• Pass: correct model, --video used, text metrics summarized with business interpretation, and the Open report URL included.
• Partial: correct model but treats the result URL as the only output.
• Fail: uses a video generation model, asks for a generation prompt, or ignores the uploaded clip.

Scenario 12 — Video explainer preset selection

User request:

Make a one-minute narrated explainer in Russian about how neural networks learn. Show me the available visual styles first.

Expected behavior:

• Routes to higgsfield-video-explainer, not generic video generation.
• Runs higgsfield preset list video-explainer --json and shows current preview links.
• Waits for style selection, then resolves its UUID with higgsfield preset resolve video-explainer.
• Lets the user choose one voice and creates six Russian seed_audio narration jobs first.
• Creates six matching 10-second gemini_omni clips second, all with the resolved style key.
• Builds ordered blocks.json pairs and finishes with one explainer_video generation job.

Score:

• Pass: live preset and voice selection, six matched audio/video blocks, final assembled MP4.
• Partial: correct block pipeline but chooses a requested option without showing it or loses a block pairing.
• Fail: embeds a preset list, fabricates an ID, uses video_explainer, or returns loose clips.

Scenario 13 — Playable browser game

User request:

Build a small mobile-friendly browser game where a fox collects stars, then give me a play link.

Expected behavior:

• Routes to higgsfield-game-generation.
• Produces a STYLE FORMULA and design/assets.csv before visuals/code.
• Builds responsive touch + keyboard controls and a complete win/lose/restart loop.
• Verifies locally, packages the required root layout, and runs higgsfield game deploy.
• Does not run higgsfield game publish unless marketplace publication is requested.

Score:

• Pass: coherent game, manifest-driven assets, local verification, returned deploy URL.
• Partial: playable and deployed but missing one planning artifact or mobile verification.
• Fail: returns only concept art/code, constructs a URL manually, or publishes publicly without consent.

Scenario 14 — Complete visual identity (brandkit)

User request:

Create a visual identity and brandbook for Northline, a calm technical logistics platform. Start from scratch.

Expected behavior:

• Routes to higgsfield-brandkit, not generic generation or Marketing Studio's URL-import brand kit.
• Asks only for unresolved context/preferences and checks stage-specific local dependencies without installing them silently.
• Creates 2–3 palette boards, shows them, and waits for a selection.
• After palette approval, generates exactly three recraft_v4_1 vector SVG marks with identical technical parameters and waits for logo selection.
• Proposes typography only after logo selection, persists each approval in project-local brandkit/state.json, and builds the canonical PPTX/PDF Brandbook only after all three slots are approved.
• Does not use FFmpeg, self-select a logo, invent brand claims, or replace an approved mark during downstream mockups.

Score:

• Pass: correct staged routing, three SVG candidates, explicit approval gates, durable state, and dependency-aware Brandbook output.
• Partial: correct assets but asks a full questionnaire, installs tools without permission, or loses approval state.
• Fail: uses the Marketing Studio metadata kit, self-approves a mark, produces only raster logos, or skips palette/logo/type selection.

Round template (copy when recording results)

Round: <N>
Date: <YYYY-MM-DD>
Commit: <sha>
Skills version: <0.x.x>

Scenario 1: pass | partial | fail — <one-line reason>
Scenario 2: ...
...
Scenario 10: ...
Scenario 11: ...
Scenario 12: ...
Scenario 13: ...
Scenario 14: ...

Aggregate: <P pass / Q partial / F fail>
Time-to-result mean: <Ns>
Notable regressions: <list>