Generate YouTube Thumbnail

A reusable workflow for creating YouTube thumbnails via Novoads' POST /v1/images with nano-banana-pro, with proper character likeness and proven CTR formulas.

When to use this skill

Trigger on phrases like:

• "make me a YouTube thumbnail"
• "create a thumbnail for this video"
• "I need thumbnail variations / A/B tests"
• "remake this thumbnail with my face"
• "generate 10 thumbnail concepts"
• "thumbnail with [me / my product / my brand]"

Read order

Paths below are relative to this skill's own folder, and every file they name ships inside it. That holds wherever the folder lands: the repo, the .claude/skills/ and .cursor/skills/ copies sync-skill.sh makes, or a standalone install of just this skill.

1. This file — workflow, decision tree, batch generation
2. prompting/guide.md — likeness alignment, expressions cheat sheet, prompt structure
3. prompting/formulas.md — 5 proven thumbnail formulas with templates
4. scripts/generate-batch.sh — the batch script (upload once → generate → download)

For the API contract itself — error codes, rate limits, the upload flow — the novoads-api skill's reference.md is the authority.

Prerequisites

• .env with NOVOADS_API_KEY (novo_ + 64 hex). Verify with ./scripts/check-novoads-env.sh.
• Optional PRODUCT_ID. Omit it and jobs land in your default product; productId is organizational only and does not influence what is generated.
• Reference images on disk (NOT pasted in chat — chat-pasted images are NOT accessible to the API):
  • face/ — several photos of the subject (headshot + 3/4 angles + close-ups + expressions)
  • logos/ — brand logos as files
  • products/ — clean product shots
  • examples/ — real ad screenshots, comparison material
  • style/ — example thumbnails the user wants to match aesthetically

REST key required. A Novoads MCP connector is not a substitute. If NOVOADS_API_KEY is missing or still the placeholder, stop before any generation work and tell the user: "Before continuing, create an API key at https://novoads.ai/dashboard/settings?tab=api and paste it into .env." That holds even when mcp__novoads__* tools are connected and authenticated in the session. Never call mcp__novoads__* tools from this repo's workflows: they are a different surface with different behavior, including the units they quote costs in. Repo installs verify with ./scripts/check-novoads-env.sh; a solo install checks NOVOADS_API_KEY in the environment.

Pack version. Every /v1 response carries X-Novoads-Pack-Version; mention a newer pack at https://github.com/novoads/agent-skills only when that header names a version NEWER than this file's metadata.packVersion — equal or older is nothing to say, and it is never a reason to stop.

If references are missing or the user pastes images in chat instead of saving them, stop and ask the user to drop the actual files into a project folder (e.g. references/youtube thumbnail/). Chat paste ≠ file on disk.

References — how many, and why it changed back

nano-banana-pro accepts up to 14 referenceAssetIds per generation, which is fal's own published ceiling for the /edit arm this workflow calls, confirmed by probe on 2026-08-03. This skill runs on nano-banana-pro, so the fork's original advice applies here unchanged.

For a while this pack said the cap was 4 on every image model and that "always use 5+ face references" was void. That was our own held number, carried from a shared constant and never tested against the transport we serve this model from. It has been corrected.

Two facts still shape how you spend them:

1. An assetId is durable and reusable. Upload once, reference forever, across runs and across sessions. Uploading is free and unlimited — only generations are charged. So keep a library of uploaded face angles and brand assets and pick from it.
2. The other image models are lower. reve-2.1 takes 8; gpt-image-2 takes 4, and that 4 is measured, not assumed — a 5-reference body was refused on the same 2026-08-04 probe run. If you route a likeness run to gpt-image-2, the 5+ advice does not survive the trip; say so rather than silently citing four.

Always use 5+ face references for character work. With 1-2 the model generalizes to a generic face; with 5+ from different angles it locks in the specific person. Spend them: headshot, 3/4 angle, studio close-up, one smiling, one neutral, plus any brand asset that must be pixel-accurate.

Rolling window for a series (ported from the character-sheet workflow): when generating variation N, pass [hero, N-1, N-2, N-3, N-4] — the approved hero shot as the anchor plus the four most recent good outputs. Drop any variation that drifted rather than feeding it forward; a bad reference propagates.

If the user insists on more coverage, the answer is more prompt specificity, not more refs.

Workflow

1. Gather requirements (in order)

Ask the user for any missing context, but only what you actually need:

1. Concept — what's the video about? Single concept, A/B variations, or specific recreation of an existing thumbnail style?
2. Subject — who is in the thumbnail (the user themselves, an AI character, no person)?
3. Brand assets — which logos / products / brand colors should appear?
4. Text — what should the title text say? Will text be baked in, or added in post (Canva/Photoshop)?
5. Comparison material — for "real vs AI" thumbnails, what real ad and what AI-generated ad?

2. Verify references exist on disk

ls "references/youtube thumbnail/"

If references are missing, ask the user to drop them. Do not proceed with text-only descriptions for brand-specific items (logos, branded products, branded apparel) — you'll get generic AI approximations that don't match the brand. Generic descriptions are OK for backgrounds, expressions, and clothing.

Then pick the ones that will be cited — 5+ face angles for character work, up to 14 in all (see the cap section above).

3. Price the batch and confirm (MANDATORY)

curl -sS -X POST "$NOVOADS_BASE_URL/v1/estimates" \
  -H "Authorization: Bearer $NOVOADS_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"kind":"image","model":"nano-banana-pro","prompt":"<one composed prompt>","numImages":1}'

It charges nothing, and it is the only legitimate source of a price. Multiply by the number of variations and show the user the total against their balance before firing anything. Wait for an explicit yes.

Never quote a credit figure from memory, from logs/novoads-api.jsonl, or from MASTER_CONTEXT.md — there are no credit numbers written down anywhere in this repo, on purpose. Name the model in the estimate body: the image models' schedules differ by more than 3×, and an estimate that omits model prices gpt-image-2.5-sunburst, understating a nano-banana-pro batch.

The call says nothing about the prompt — no endpoint on this API reads one for quality. Re-read the prompt against the formula yourself before pricing: a batch runs the same shape N times, so a flaw in it is paid for N times.

4. Pick a formula

See prompting/formulas.md for the 5 proven formulas. Match the user's intent:

User says...	Use formula
"Just me with my brand" / "branding thumbnail"	Peace-sign / branding
"Real vs AI" / "compare" / "before/after"	Real vs AI comparison
"Show the process" / "with the terminal"	Terminal flow
"Surprised face" / "shocked reaction"	Reaction shock
"Replace" / "alternative" / "swap out"	Before/after split

5. Compose prompts

Follow the template in prompting/guide.md:

YouTube thumbnail, 16:9 landscape.
[SUBJECT — likeness block + clothing + framing + "no hands" if applicable]
Expression: [specific expression from expressions cheat sheet]
[LEFT visual element + reference]
[RIGHT visual element + reference]
Across the top in massive bold yellow block letters with thick black outline reads [TITLE].
Background: [color + glow]
Style: [aesthetic notes]
Avoid: distorted face, extra fingers, hands visible, blurry logos, generic face

Always include the CRITICAL CHARACTER LIKENESS block when the subject is a real person. It and the references do different jobs: the refs fix what the face IS, the block fixes what the prompt must not let drift. Neither substitutes for the other, so write both.

Prompts are capped per generation; the API names the limit if you exceed it. Reference the uploaded images positionally ("the face in the first reference image") — order is preserved.

6. Generate (use the batch script)

Copy scripts/generate-batch.sh to a new versioned script (scripts/generate-thumbnails-vN.sh) and modify:

1. Update REF_BASE and COMMON_REFS (max 14 — the script's REF_CAP) with your reference file paths
2. Replace the PROMPTS array entries with your composed prompts
3. Run with bash scripts/generate-thumbnails-vN.sh > output/run.log 2>&1 &
4. Monitor with tail -F output/run.log | grep -E "DONE|FAILED|charged"

The script handles:

• Image preprocessing (Lanczos to 1080px longest side, RGB JPEG)
• Upload once, reuse the assetIds for every prompt in the batch
• Each call blocks ~60–90s and returns the image; one still going after about 105 s answers running and the script polls it (API 2.39.0)
• Bounded parallelism (4 in flight, under the API's concurrency ceiling of 5)
• Per-run credit total from each response's creditsCharged

7. Review and present

After all generations complete, read each thumbnail with the Read tool and present:

• Brief verdict per thumbnail (likeness, readability, emotional impact)
• Top 3 picks ranked by CTR potential
• Specific reasons for the picks (which expression, which color contrast, which formula)
• Offer next-step refinements (different expression, background color, copy variation)
• The total credits actually charged, summed from the responses — not the estimate

Quirks and pitfalls

Reference assetIds are durable — reuse them

This is the opposite of how the previous backend worked, where a reused reference caused HTTP 500 UNKNOWN_ERROR and every generation needed a fresh upload. On Novoads, POST /v1/uploads returns an assetId that keeps working across calls and across sessions. Upload once per batch — or once per campaign — and reuse. Re-uploading the same face photo produces a different id and quietly loses the anchor that was holding likeness steady.

The presigned uploadUrl does expire (expiresInSeconds in the response), but that only affects the PUT window, not the assetId.

The signed upload headers must be echoed exactly

POST /v1/uploads returns a headers object. The PUT must send those headers byte for byte — Content-Type and Content-Length are both part of the URL signature, so storage answers 403 if either differs, including an added ; charset=…. Don't let an HTTP client infer the type.

Image preprocessing is practice, not a documented requirement

The old "images smaller than 1080px longest side return 422 — image too small" rule was specific to the previous backend and has not been verified against Novoads. The batch script still upscales, because small references genuinely produce worse likeness — but do not report the 422 claim to a user as an API rule. (Flagged for confirmation in the Phase 4 live smoke.)

referenceAssetIds is an array of plain strings

Not objects, and not file paths — the assetId values returned by POST /v1/uploads. Maximum 14 on nano-banana-pro, which is the model this skill runs on. The other image models are lower and the number does not travel: reve-2.1 takes 8, gpt-image-2 takes 4.

Chat-pasted images are NOT files

If the user pastes an image directly in chat, you cannot pass it to the API. Ask them to save the actual file into a project folder.

Likeness drift with too few references

With 1-2 face references the AI generalizes to "generic bearded man with glasses." With 5+ face references from different angles it locks in the specific person. Always use 5+ face references for character work. For a series, use the rolling window described above.

Concurrency is 5, not "as many as you like"

The API allows 5 concurrent generations per organization. A 6th in flight returns 429 with details.reason: concurrency_limit — and unlike the other 429 causes, slowing your request rate does not help; only a finishing job frees a slot. The batch script caps itself at 4 in flight for headroom.

Do not blindly retry a failed generation

There are no idempotency keys on this API, so resubmitting after a timeout or a 500 can render and charge twice. Check GET /v1/generations for a job that already landed before resubmitting. The batch script deliberately does not auto-retry.

macOS bash 3.2

Default macOS bash doesn't support declare -A (associative arrays). The batch script uses indexed arrays instead.

Brand-specific items need actual reference files

Text descriptions of brand-specific items (logos, branded apparel, custom merchandise) produce generic approximations. For pixel-accurate brand reproduction, save the actual brand asset to disk and cite it as one of your references.

Cost

There is no cost table in this file, deliberately. Every price comes from a live POST /v1/estimates in the current session, shown to the user and approved before anything is generated (see step 3). Report the real total from creditsCharged when the run finishes.

nano-banana-pro costs more per image than gpt-image-2. If a batch is exploratory and the concept doesn't depend on photoreal likeness, pricing both is worth the two estimates, which charge nothing.

See also

• prompting/guide.md — likeness alignment, expressions, prompt structure
• prompting/formulas.md — 5 proven CTR formulas with prompt templates
• scripts/generate-batch.sh — the batch generator
• the novoads-api skill — endpoint, upload flow, error codes, rate limits
• the nano-banana-image-ad skill — the same model, pointed at Meta ad creatives