Iverywhere

AI Asset Playbook

The strategy and the recipes for AI-assisted assets: which tools, which workflows, and the exact starter prompts — built so real photography stays the autobiography and AI plays everything else. Created July 8, 2026.

How this plugs in. This is the tools-and-workflows layer under three existing laws: the art-direction bible (house grade, the six city AI recipes, "never let AI invent the product"), the b-roll library (the three absolutes; AI = atmosphere lane only), and the content calendar (where every asset lands). Nothing here overrides those; this doc answers which tool, which buttons, which prompt.
The strategy

One page: what AI does and never does

Real photography carries the brand. AI multiplies it. Every asset lives in exactly one of four lanes — the lane decides the tool, the rules, and whether it's logged as AI.

LaneWhat it isToolsDisclosure & log
1 · RealYour photos and phone b-roll, graded to the house preset. The autobiography — pairing frames, family, product on real people.Camera + Lightroom (house preset)None. Not AI.
2 · AssistedReal frame, AI-touched: extend a graded photo to 9:16, clean a distraction out of a background, relight a corner. The photograph is still the photograph.Photoshop generative expand / fill (Firefly)Log it. No public label needed for cleanup-level edits; never "assist" a face, a body, or the garment.
3 · CompositeReal product pixels, generated world. A real still-life or flat-lay photo of the vest, environment built around it. Product is untouched — that's the law.Nano Banana Pro (Gemini), PhotoshopLog it + label when published as imagery of the product in a place ("studio composite" energy — never captioned as a real trip).
4 · AtmosphereFully synthetic place-and-texture frames and clips. No people, no product, no autobiography. The Como shoreline behind a planning-season Reel.Midjourney / Nano Banana Pro → Veo (motion)Log it + expect platform auto-labels (see compliance). Caption never claims "our trip."
The three absolutes still govern (from the b-roll doc): (1) never AI-generate the product; (2) never AI-generate people presented as real — no synthetic families, ever; (3) AI plays atmosphere, not autobiography. The one extension this doc adds: never animate the garment with AI either — image-to-video models warp fabric, quilting, and logos; motion belongs to the environment (see Workflow 4).
Compliance, mid-2026

The disclosure reality — why the log is non-negotiable

The stack

Tools — two subscriptions cover 90%

Chosen for replicability and simplicity, not maximal capability: fewest tools that hold the aesthetic, all driven by prompt templates you reuse. Roughly $40–60/mo incremental on top of Adobe you already own.

JobPrimaryWhy this oneBackup / when to switch
Image — edit & compositeNano Banana Pro (in Gemini, ~$20/mo Google AI Pro)Best-in-class at instruction-following edits and multi-image reference consistency: feed it the real product photo + a reference frame of the house look and tell it what to keep. This is the Lane 3 engine.Photoshop generative fill for surgical local edits where you need layer control.
Image — atmosphereMidjourney v7 (~$10–30/mo, optional)Still the best raw aesthetic for filmic, painterly atmosphere plates (Lane 4). Weak at edits — use it for first-generation beauty only.Nano Banana Pro does this well enough to skip the subscription; add MJ only if atmosphere volume grows.
Image — extend & cleanPhotoshop / Firefly generative expand (owned)Adobe-stock-trained = the commercially safest generator; integrated with the Lightroom grade workflow. The Lane 2 engine: one hero frame → 1:1, 4:5, 9:16, 16:9.Lightroom Remove for small distractions.
Video — image-to-videoVeo 3.1 (in the same Google AI Pro sub, via Gemini/Flow)Strongest all-rounder in mid-2026 — best fabric/light physics, native audio, vertical output; its default look is closest to high-end stock, which after the house grade reads editorial. Start every video from a graded still, never from text alone.Kling 3.0 for volume on a budget (~$1/10s clip). Runway Gen-4.5 when you need motion-brush control (animate background only, lock the product). Do not build on Sora — OpenAI is discontinuing it (app Apr 2026, API Sept 2026).
MusicNative Instagram audio first; ElevenLabs Music (~$5–22/mo) for owned surfacesReels should ride licensed in-app audio (reach + zero risk). For the site, email GIFs→video, YouTube, and anything embedded, ElevenLabs has the cleanest commercial license in the market — licensed training data, explicit ad/commercial rights, no litigation history.Suno Pro also grants commercial rights but sits under active litigation and distributor blocks — fine for drafts, not for the brand's permanent surfaces.
Voice / speechReal founder voice + Adobe Enhance SpeechThe voice doc's whole thesis is that Haley and Trevor are the trust signal — VO stays human. AI's job is cleanup: kitchen-table recordings in, studio-adjacent audio out.ElevenLabs TTS only for internal scratch tracks while editing; never published.
Virtual try-on / AI modelsNone — deliberately.The 2026 fashion-AI category (Botika, FASHN, Claid et al.) exists to put garments on synthetic models. That's absolute #2 broken as a service. Coordinated real families are the brand's one selling point; renting fake ones would spend the moat to save a shoot.Revisit only for fit-visualization R&D, never for published imagery.
Workflow 1 · Lane 2

Extend a real frame to every aspect ratio

The highest-ROI, lowest-risk workflow: one hero photograph from the July 24 shoot becomes site hero (16:9), feed (4:5), story (9:16), and Pinterest (2:3) without cropping away the negative space that is the brand.

  1. Start from the graded keeper (house preset already applied in Lightroom).
  2. Photoshop → Crop tool set to the target ratio, expand the canvas on the side with sky/water/wall — extend the emptiness, never the subject side.
  3. Generative Expand with an empty prompt first (it usually continues sky/stone/water cleanly). Only add words if the fill drifts: continuation of overcast sky, same soft light.
  4. Zoom 200% along the seam — check for texture repeats and tone jumps. Regenerate rather than heal.
  5. Export → log as Lane 2 → file as iv-ai-extend-<scene>-<ratio>-<nn>.
Rule of thumb: AI may add up to ~40% new canvas of emptiness (sky, water, wall, pavement). If the expand would need to invent architecture, people, or product, stop — that's Lane 3 or a re-crop.
Workflow 2 · Lane 3

Real product, generated world

The composite recipe from the art-direction doc, made concrete. Input: a clean, graded photo of the real garment (flat-lay, hung, or still-life — the A6/P2/P5 frames you're already capturing). Engine: Nano Banana Pro with the product photo as a hard reference.

  1. Shoot or pick the real product frame: even light, minimal background, garment filling ~30–50% of frame. Grade it first.
  2. In Gemini, attach the product photo (and optionally one on-brand reference frame for the look) and use the template below. The instruction that does the work: the garment pixels are untouchable.
  3. Generate 4; pick on environment quality only — any candidate that altered the garment is dead on arrival regardless of how good the world looks.
  4. Run the garment QC (below) at 200%. Then re-grade the whole composite lightly so generated world and real product share one treatment.
  5. Log as Lane 3; caption it as what it is (a still-life, product in a place) — never as trip autobiography.
Using the attached photograph, keep the garment EXACTLY as photographed — identical
fabric weave, quilting lines, stitching, zipper, drape, color, and label. Do not
regenerate, redraw, sharpen, or "improve" any pixel of the garment.

Replace only the environment: place the vest [draped over a rattan trattoria chair
on a stone terrace before opening / folded on a weathered teak bench beside a still
grey-green canal / laid on ivory linen in a villa window's soft north light].

Light the scene to match the garment photo's existing light direction and softness.
Muted, film-like editorial mood: lifted matte blacks, soft fine grain, restrained
palette of Brunswick green, ivory, stone, and soft sage. Vast quiet negative space;
the garment sits low and off-center. No people. 4:5 vertical.
Avoid: altered garment, added logos or text, warm golden push unless specified, harsh
shadows, busy props, HDR gloss, perfect symmetry, catalog-studio look.

Garment QC — 60 seconds, every composite

Workflow 3 · Lane 4

Atmosphere plates — the six city recipes, operationalized

The art-direction doc's six recipes are the prompt library; this is how to run them. No people, no product — those clauses stay in every prompt verbatim.

  1. Pick the recipe for the calendar beat (Copenhagen / Amsterdam / Watson's Bay / Lisbon / rainy morning / still-life — or the b-roll doc's six scene skeletons for shorter prompts).
  2. Generate at 4:5 and 9:16 in the same session (same prompt, two ratios) so feed and story versions match. In Midjourney add --ar 4:5 --style raw; in Gemini just state the ratio.
  3. Batch of 4–8; select with the b-roll filter (palette · light · texture · motion · Sézane test).
  4. Grade the keeper with the house preset — non-negotiable; this is what makes an AI frame sit inside the brand rather than beside it. Add grain last (grain amount 12 / size 20) — it also buries the residual "AI smoothness."
  5. Log as Lane 4 → assets/art-direction/broll/ai/iv-ai-atmo-<scene>-<nn>.
Prompt anatomy — every atmosphere prompt has five blocks, in order: (1) place + one concrete anchor object · (2) light, true to the place (warm OR cool — never default-golden) · (3) texture callout (linen, stone, water, straw) · (4) the grade sentence: "film-like editorial grade, lifted matte blacks, soft fine grain, muted saturation, palette of Brunswick green, ivory, stone, soft sage" · (5) the negatives: no people, no product, no logos, no text, no HDR, no oversaturation. When iterating, change exactly one block per generation — that's what makes results reproducible instead of a slot machine.
Workflow 4 · Lanes 2–4

Still → quiet motion (the Reel b-roll engine)

The single biggest content multiplier: a graded still (real or AI) becomes 6–8 seconds of ambient video for Reels. Image-to-video only — never text-to-video — so the frame you approved is the frame that moves. Veo 3.1 via Gemini/Flow; the motion rules from the art-direction doc apply to AI motion exactly: slow, quiet, nothing bounces.

  1. Start from a graded, approved still. (Real photo in = Lane 2-ish "living photo"; AI plate in = Lane 4. Both logged.)
  2. Prompt motion in the environment only — light, water, fabric-of-the-place (curtains, tablecloths), haze. Camera either locked or a barely-perceptible push.
  3. Generate 8s vertical (9:16) for Reels; 4–6s is often enough — better to loop a short perfect clip.
  4. Watch every frame at full size: i2v models drift — a chair leg bends, water develops a physics glitch, and if any garment is in frame, fabric morphs. One bad frame = regenerate.
  5. Mute the native audio by default (use in-app licensed audio on Reels). Log, then file under broll/ai/.

Starter motion prompts (paired to existing scene recipes)

[Como morning plate] Camera locked. Morning haze drifts almost imperceptibly across
the water; tiny glints move on the surface; nothing else changes. Slow, quiet,
documentary stillness. No people, no boats moving, no camera shake, no zoom.
[Empty trattoria terrace plate] Static camera. Dappled light through the vine canopy
shifts gently as if a slow breeze; one linen tablecloth corner lifts once, softly,
and settles. Everything else still. No people, no birds, no fast motion.
[Riviera pool edge plate] Locked frame. Water caustics ripple slowly across the
travertine; the out-of-focus umbrella fringe sways a few millimeters. Heavy, warm,
unhurried afternoon. No people, no splashes, no camera movement.
[Real packing-table photo] Barely perceptible 2–3% push-in over 8 seconds; window
light intensity breathes very slightly as if a cloud passes. NOTHING in the frame
moves or deforms — no fabric motion, no object motion. Preserve every detail of the
photograph exactly.
Product-in-frame rule for video: if the garment is visible, the only permitted motion is camera motion (slow push/parallax) or light — never fabric or object motion. If the vest so much as ripples, the quilting will smear. When you need environment motion around a product, generate the plate without the product (Workflow 3), animate it (this workflow), and keep the product to stills.
Workflow 5 · Sound

Audio — mostly restraint

Music

  • Reels/Stories: in-app licensed audio, chosen quiet (original audio also gets a ranking boost — a graded clip with its own room tone is on-brand and favored).
  • Owned surfaces (site video, YouTube, lead-magnet pages): ElevenLabs Music, one reusable "house bed" — generate once, reuse everywhere, exactly like the photo grade.
  • Never: AI covers, artist-style prompts ("like Norah Jones"), or Suno tracks on permanent brand surfaces while its litigation runs.
Quiet instrumental, 70 bpm. Warm nylon-string guitar and soft felt piano, close
and roomy like a living-room recording. No drums, no builds, no melody hook —
texture, not song. Slightly tape-worn, ends unresolved. 60 seconds, loopable.

Voice

  • VO is always Haley or Trevor, recorded on a phone in a soft room (closet, car, duvet fort).
  • Clean it with Adobe Enhance Speech — kitchen-table in, studio-adjacent out. That's the entire AI role.
  • No voice cloning on anything published. Scratch TTS for timing an edit is fine; it never ships.
  • Script check: read it aloud once — if it sounds like copy, rewrite it to sound like the group text (voice doc, rule 9).
Replicability

The prompt library system

The aesthetic lives in reusable blocks, not in heroic one-off prompts. Treat prompts like the Lightroom preset: master blocks stay fixed; one variable changes per run.

ElementConvention
LibraryOne doc (this one + the art-direction recipes) is canonical. New winning prompts get added here with the asset they produced — a prompt without its output frame teaches nothing.
The grade sentencefilm-like editorial grade, lifted matte blacks, soft fine grain, muted saturation, palette of Brunswick green, ivory, stone, soft sage — appears verbatim in every image prompt. It's the brand's textual preset.
The negative blockno people, no product, no logos, no text, no watermark, no HDR, no oversaturation, no golden-hour default, no stock gloss — verbatim in every Lane 4 prompt.
Iteration protocolChange one block per regeneration. Keep the losers' prompts for a week (they document what doesn't work). In Midjourney, reuse --seed when refining a near-miss.
Reference imagesKeep a folder of 5 canonical "this is the look" frames (graded real photos). Attach one to every Nano Banana Pro session as the style anchor — reference beats adjectives.
Files & logiv-ai-<workflow>-<scene>-<nn> into broll/ai/; every published asset gets its capture-log row with the lane number. Ten seconds of logging = the whole compliance story.
The gate

Nothing publishes without passing

Start here

The first batch — this week

Six assets to prove each workflow end-to-end, sized to feed the pre-launch calendar's August needs. Generate, then we iterate on feedback together.

#AssetWorkflowFeeds
1Best existing graded photo extended to 9:16 + 16:9W1 (Photoshop expand)Story + site hero test
2Vest-on-chair trattoria composite from a real flat-lay/still-life frameW2 (Nano Banana Pro)Feed, PDP support
3Como-morning atmosphere plate, 4:5 + 9:16 pairW3Planning-season Reels backdrop
4That same plate animated — 8s quiet motion loopW4 (Veo 3.1)Reel b-roll bank
5One real b-roll still (packing table) as a "living photo"W4, camera-only motionReel opener
6The 60-second house music bedW5 (ElevenLabs)Site + YouTube + lead magnet
What I can run in-session: the connected Adobe tools let Claude execute parts of this pipeline directly — batch-grading to the house look, generative expand, background removal, crops to platform ratios, and Enhance Speech on VO recordings. Bring the raw frames; the assisted lane can happen in this terminal.
Appendix

Sources (landscape research, July 2026)