Multi-slide bilingual brand carousels — Claude writes the deck, kun renders Anthropic-styled slides at exact platform sizes, a human approves, channels receive

2026-07-130

social

databayt/kun

Draft, stage, and publish brand social posts — Claude drafts, /higgs renders, a human approves, Hermes relays

2026-07-130

convert

databayt/kun

Convert a file or URL to Markdown via MarkItDown (PDF, Office, images, audio, web)

2026-07-100

feature

databayt/kun

Full pipeline — idea to production (chains every stage)

2026-07-100

databayt/kun

Autonomous block QA — detect, adversarially verify, fix safe tiers, hand the residual to a human

2026-07-100

spec

databayt/kun

Technical spec — data model, file plan, refined acceptance criteria

2026-07-100

name	higgs
description	Generate and edit photos and videos for databayt org marketing, ads, and prompts using Higgsfield AI
when_to_use	Use when generating or editing marketing media — image, photo, video, ad, banner, og image, social card, hero shot, product shot, promo reel, story, avatar ad — or when a reference video/image URL needs downloading and tweaking through Higgsfield. Triggers on: /higgs, generate video, generate image, generate photo, make an ad, promo video, og image, social card, hero image, product shot, brand kit, edit video from url, download and tweak video, higgs.
argument-hint	[recipe\|prompt] [--url <ref>] [--count N] [--premium] [--ratio 16:9]

Higgs — Higgsfield media generation for databayt

One-command marketing media. Every decision below is pre-made — do NOT ask Abdout about model, style, aspect ratio, or output location. Pick the recipe, run it, deliver the file. CLI: higgsfield (aliases higgs, hf), authed via auth login (OAuth token on disk).

Zero-question doctrine

Never ask model/style/ratio — the tables below decide. Only stop for: credits insufficient for the asked job, or a truly ambiguous deliverable.
No text inside generated visuals (AI typography breaks, Arabic doubly so). Generate text-free; overlay copy in post. Exception: dtc-ads — its backend does typography.
Estimate free, then spend: generate cost <model> --prompt "...", --cost-only (dtc-ads), --enhance-only (product-photoshoot). Report spend + balance after every run.
Batch in one command — --count 1-10 (photoshoot), --batch-size 1-20 (dtc-ads) — never N separate invocations.
Always --json, parse .[].result_url, download to ~/Downloads/higgs/, deliver via SendUserFile. Job page fallback: https://higgsfield.ai/create/image
Video is expensive (7.5–22.5 cr) — iterate composition as a cheap image first (z_image, 0.15 cr), then animate the winning frame via --start-image.

Account (verified 2026-07-06)

Fact	Value
Account	`osmanabdout@hotmail.com` — free plan
Workspace	`1d67d842-bc70-40d8-8c03-576cb1b141df` (Private)
Brand kit (databayt.org)	`b0a3f528-0397-41f7-85be-0b44a458a01f`
Web product (databayt.org)	`6933f3b2-5611-4548-b015-2c963d054dc4`
Output dir	`~/Downloads/higgs/YYYY-MM-DD-<slug>.<ext>`

Preflight (one call, fix only what fails)

higgsfield account status || { higgsfield auth login; higgsfield workspace set 1d67d842-bc70-40d8-8c03-576cb1b141df; }

auth login opens browser OAuth (default browser holds the higgsfield.ai session — completes hands-free). MCP https://mcp.higgsfield.ai/mcp is registered but unauthenticated — the CLI is the primary lane; don't detour through MCP.

Model defaults (costs verified per 1 job)

Naming trap: nano_banana_2 = Nano Banana Pro; nano_banana_flash = Nano Banana 2.

Use	Model	Cost	Notes
Image draft / iteration	`z_image`	0.15	ratios 1:1,4:3,3:4,16:9,9:16
Image final	`nano_banana_flash`	1.5	+21:9, 2:3, 4:5…; `--resolution 1k/2k/4k`; image refs
Image hero / 4k / 21:9	`nano_banana_2` (Pro)	2	same params as flash
Video default	`kling3_0_turbo`	7.5	5s 720p 16:9/9:16/1:1; `--start-image`
Video from reference video	`seedance_2_0_mini`	12.5	only seedance takes `--video-references`; 480/720p
Video premium (`--premium`)	`seedance_2_0`	22.5	4k, audio, genre, ≤9 img+3 vid+3 audio refs (12 total)

Refresh after Higgsfield ships new models: higgsfield model list, params via higgsfield model get <job_type>.

Brand style blocks (append to every prompt)

minimal (default: og, hero, product, docs): "minimal premium developer-tool aesthetic, deep charcoal background, generous negative space, soft studio light, subtle film grain, monochrome with one restrained cyan accent, no text"
cinematic (default: ads, reels, promos): "high-end glassmorphic panels, deep dark mode backdrop, electric cyan and warm sunset-orange rim light, shallow depth of field, slow tracking shot, cinematic studio lighting, photorealistic, no text"

Recipes (copy-paste; swap prompt subject)

# Draft grid — explore 4 compositions for ~0.6 cr
for r in 1:1 16:9 9:16 4:3; do higgsfield generate create z_image --prompt "<subject>, <minimal>" --aspect_ratio $r --wait --json; done

# OG image / hero banner (final)
higgsfield generate create nano_banana_flash --prompt "<subject>, <minimal>" --aspect_ratio 16:9 --resolution 2k --wait --json

# Social square / story
higgsfield generate create nano_banana_flash --prompt "<subject>, <cinematic>" --aspect_ratio 1:1 --wait --json   # story: 9:16

# Promo clip (5s, default video)
higgsfield generate create kling3_0_turbo --prompt "<scene>, <cinematic>" --aspect_ratio 16:9 --wait --wait-timeout 20m --json

# Animate a winning image (image → video)
higgsfield generate create kling3_0_turbo --prompt "<motion direction>" --start-image ~/Downloads/higgs/<img>.png --wait --wait-timeout 20m --json

# Tweak a video from URL (helper: download → upload → restyle)
bash ~/.claude/scripts/higgs-tweak.sh --url "<video_url>" --prompt "<restyle>, <cinematic>"   # --premium for seedance_2_0

# Reframe existing video to another ratio (e.g. landscape → reel)
higgsfield generate workflow reframe --video <path> --aspect-ratio 9:16 --resolution 720p --wait --json

# Branded ad set — brand kit does typography/colors (batch up to 20)
higgsfield marketing-studio dtc-ads generate --prompt "<offer/hook>" \
  --format-id 18e9f327-b667-40f1-84d1-f234c67a4929 --brand-kit-id b0a3f528-0397-41f7-85be-0b44a458a01f \
  --batch-size 4 --cost-only   # rerun without --cost-only to spend
# formats: `ms ad-formats list` (Headline, Special Offer, Key Features, Social Proof, …)

# Product photoshoot — backend prompt enhancement, modes:
# restyle, social_carousel, ad_creative_pack, virtual_model_tryout, conceptual_product,
# product_shot, lifestyle_scene, closeup_product_with_person, moodboard_pin, hero_banner
higgsfield product-photoshoot create --mode hero_banner --prompt "<intent>" --image <ref.png> --count 3 --enhance-only   # preview free, drop flag to spend

Result handling (verified JSON shape)

generate create --wait --json returns an array of jobs: [{"id", "status": "completed", "result_url": "<full-res png/mp4>", "min_result_url": "<preview webp>", ...}]

higgsfield generate create <model> --prompt "..." --wait --json \
  | jq -r '.[].result_url' \
  | while read -r u; do curl -sL -o ~/Downloads/higgs/$(date +%F)-<slug>-$RANDOM.${u##*.} "$u"; done

Async batch (fire many, wait once): capture .[].id per create (no --wait), then higgsfield generate wait <id> --json. List recent: higgsfield generate list.

Budget ladder (free plan ≈ 10 cr — treat credits as scarce)

Drafts are ~free (66 z_images / 10 cr); video ≥ 7.5 cr — confirm before multi-video runs.
If a job costs more than the remaining balance, stop and report: asset plan, per-job cost, balance, and that the workspace needs an upgrade (https://higgsfield.ai/pricing) — a billing change, so it's Abdout's call (subscription doctrine).
higgsfield account transactions --size 20 audits spend.

Recovery

Symptom	Fix
`Not authenticated`	`higgsfield auth login` (browser OAuth, hands-free)
`No workspace selected`	`higgsfield workspace set 1d67d842-bc70-40d8-8c03-576cb1b141df`
Unknown model / param error	`higgsfield model list`, `higgsfield model get <job_type>`
Job stuck	`higgsfield generate get <id> --json`; re-wait with `generate wait <id>`
Brand kit missing	`higgsfield ms brand-kits fetch --url https://databayt.org --wait --json` (free), update ID here