[ CREATIVE ]

/krea-create-images

Generate ad images for a ACME Agency client using Krea.ai Nano Banana 2.

Download the skill file (.md)

/krea-create-images

Generates ad creatives for a ACME Agency client using Krea.ai Nano Banana 2 (nano-banana-flash). Reads brand context, builds an optimized prompt, generates images, uploads to Drive, and reports to Slack.

Trigger

• /krea-create-images [ClientName]
• /krea-create-images ACME Agency --ratio 4:5 --brief "dental clinic, happy patient"

Preflight (run BEFORE any expensive Krea call)

Validate every assumption that has burned us in production. If ANY check fails, abort with a clear message naming the failure. Do not retry, do not call diagnostician — preflight is the cheap fix path.

1. Client exists in ACME Agency/clients/clients.json (case-insensitive match). If missing → abort with the list of valid clients.
2. Ratio is valid for Krea NB2: 1:1, 4:5, 9:16, 16:9, 3:2, 2:3. Reject anything else.
3. Batch is 1-4 (Krea API max). If passed >4, clamp and warn.
4. Resolution is 1K, 2K, or 4K. Default 2K.
5. Brief OR reference URL is present. If neither → abort with usage hint.
6. Reference URLs (if passed) are publicly fetchable — they must be HTTP(S), NOT local file paths. If a local path was passed by mistake, upload it via uploadPublic() from lib/google_drive.mjs first.
7. KREA_API_KEY is set in .env. If missing → abort with "add KREA_API_KEY to .env".
8. drive_folder_id is present if --drive (default). If missing → either skip drive cleanly OR abort if explicitly required.
9. Slack channel resolves if --slack (default). If getClientChannel() returns null and bot isn't invited → skip with warning, don't crash.

Workflow

Step 1 — Identify client

• Parse the client name from the argument
• Look up the client in ACME Agency/clients/clients.json
• Confirm the match with the user if ambiguous

Step 2 — Load brand context

• Read ACME Agency/clients/<ClientName>/CLIENT.md if it exists
• Extract: products/services, target audience, brand tone, colors, market
• If CLIENT.md is missing, note it and rely on the user's brief

Step 3 — Gather brief (ask if not provided)

If no --brief was given, ask the user:

1. What's this image for? (ad creative, organic post, story, banner, etc.)
2. Any specific copy or text to include in the image? (headline, CTA, phone number, etc.) — Nano Banana 2 handles typography well
3. Preferred visual style? (photorealistic, clean minimal, lifestyle, product shot, architectural, etc.)
4. Preferred color palette? (or derive from CLIENT.md if it has brand colors)
5. Ratio? → default 4:5 for feed ads, 1:1 for square, 9:16 for stories/reels

Step 4 — Build the Krea prompt

Construct a detailed, well-structured prompt. Nano Banana 2 works best with:

Structure:

[Main subject / scene] — [style descriptor] — [lighting/mood] — [text to overlay if any] — [color palette] — [target feel]

Tips:

• Lead with the most important visual element
• Be specific: "woman in white lab coat smiling at camera, dental clinic background" beats "dentist"
• Include text overlay instructions explicitly if needed: e.g. bold Croatian text "Zakažite pregled" bottom center
• Name colors: "white, navy blue, light grey" not "brand colors"
• Include market/cultural cues: "Croatian market", "Mediterranean setting", "modern European"
• Do NOT use negative prompts — Nano Banana 2 does not support them

Example prompts:

• professional dental clinic reception, female dentist in white coat smiling warmly, clean white interior, soft natural light, modern European aesthetic, Croatian market, photorealistic
• luxury apartment building exterior, golden hour lighting, Mediterranean architecture, modern minimalist, warm tones, real estate advertisement style

Step 5 — Run the script

node ACME Agency/scripts/krea_create_images.mjs "ClientName" \
  --ratio 4:5 \
  --brief "your constructed prompt here" \
  --batch 2 \
  --resolution 2K

Default settings:

• Ratio: 4:5 (best for Meta feed ads)
• Resolution: 2K
• Batch: 1 (increase to 2–4 for variations)

Note on --batch: Krea removed the API-side batchSize field on 2026-05-26 (one POST → one image). lib/krea.mjs now fires N parallel POST requests when --batch N is set, so the public behavior is unchanged from a caller's perspective — you still get N images back. Latency is roughly the same as before (parallel) but each image counts as one Krea credit instead of N images per credit (refer to current Krea pricing for the exact cost delta).

Common ratio choices:

Automatic logo reference (always-on):

When <ClientName> is passed, the script auto-includes the client's official logo as the first reference image (via ensureClientLogo() from lib/logo_prepare.mjs). It resolves from clients.json > reference_assets.logo, then a local logo.png in the client folder, then a Playwright screenshot of the website. The prompt is also augmented with a "logo is canonical, reproduce pixel-exact" instruction. No flag — this is on by default to fix Nano Banana's tendency to redraw logos.

If the log says [Logo] No logo available, the client has no landing_page and no local logo file — drop a logo.png into ACME Agency/clients/<ClientName>/ and re-run, or run node ACME Agency/scripts/lib/logo_prepare.mjs "<ClientName>" manually.

Reference-image flags (for "make something like this" requests):

Krea NB2 can clone style and composition from a reference image. Use these flags when the user supplies an image as inspiration:

# Single reference (most common)
node ACME Agency/scripts/krea_create_images.mjs "ClientName" \
  --brief "in the style of the reference, but for a dental clinic" \
  --reference-url "https://drive.google.com/uc?id=ABC123"

# Multiple references (up to 14)
node ACME Agency/scripts/krea_create_images.mjs "ClientName" \
  --brief "fuse these visual themes" \
  --reference-urls "url1,url2,url3"

# Reference-only (no brief — useful for quick clones)
node ACME Agency/scripts/krea_create_images.mjs "ClientName" \
  --reference-url "https://drive.google.com/uc?id=ABC123"

# Style transfer with strength control
node ACME Agency/scripts/krea_create_images.mjs "ClientName" \
  --brief "modern dental clinic" \
  --style-image-url "https://drive.google.com/uc?id=XYZ" \
  --style-strength 1.5

Important: reference URLs must be publicly fetchable by Krea. The proven path is:

1. Upload the local image via uploadPublic(localPath, parentFolderId) from ACME Agency/scripts/lib/google_drive.mjs
2. The function returns a https://drive.google.com/uc?id=<fileId> URL
3. Pass that URL to --reference-url

This is the same pattern static_ads_generate.mjs already uses in production for product-image references.

When the creative-director agent invokes this skill: the agent will handle the upload-to-public step automatically. Skill scripts should never assume the user has a public URL handy — if the user gives a local file path, the upload step is part of the workflow.

Step 6 — ACME Agencyw & iterate

After generation:

1. Show the user the Drive links and local paths
2. Ask: "Want a variation, different style, different ratio, or adjusted copy?"
3. If iterating, tweak the prompt and re-run — append /v2, /v3 to track versions mentally
4. Offer to generate a batch of 2–4 with slight prompt variations for A/B testing

Step 7 — Output

The script automatically:

• Downloads images to tmp/krea_<timestamp>_<n>.jpg
• Uploads to client's Google Drive folder (via drive_folder_id in clients.json)
• Posts Drive links to the client's Slack channel
• Saves a manifest to ACME Agency/clients/<ClientName>/krea_generations.json

Verification (run AFTER Step 7 — prove it landed, don't assume)

Rubric: verify.md. Two layers — one enforced, one you run.

Deterministic (automatic, in krea_create_images.mjs). The script now runs validateFiles() from shared/verify/assert_output.mjs right after generation: 0-byte / sub-10 KB stubs are dropped before upload, and if NONE are valid it exits non-zero instead of posting an empty result. The manifest records dropped. Still confirm from the output: valid image count = batch (a short count is logged, not swallowed), driveUrls map 1:1 to the valid images, Slack returned a non-null ts.

Vision quality (you run, before reporting to the user). Invoke the verifier agent on the generated image(s) against verify.md (threshold 8): is the logo/text real and legible, does it match the brief, is it on-brand and usable as an ad? Re-run with a tweaked prompt for anything < 8 (name the defect); offer a 2–4 variation batch when style is the issue.

Never claim success while either layer shows a gap — give the specific evidence.

Key Files

API Details

• Endpoint: POST https://api.krea.ai/generate/image/google/nano-banana-flash
• Auth: KREA_API_KEY in .env
• Polling: GET https://api.krea.ai/jobs/{id} until status: completed
• Result: result.urls[] — direct CDN download URLs
• Cost: ~48 compute units per image, ~15s generation time
• No negative_prompt support on this model

Troubleshooting