Gemini Image Handoff

Gemini Image Handoff

A semi-automatic route to a Gemini image that needs no API billing. The skill does one half automatically — author a prompt that is optimised for Google's native Gemini image model — and hands the other half to the operator: paste the prompt into the Gemini web UI, generate, and download the image from the chat. It makes no API call, needs no GEMINI_API_KEY, and writes no image and no sidecar; the operator owns where the downloaded file lands.

It exists because the gemini provider of the image-generate tool requires billing (gemini-2.5-flash-image reports a Free-Tier quota of limit: 0). This skill keeps Gemini reachable for occasional use through the free chat UI, while still applying the verified model baseline so the pasted prompt is actually optimised for Gemini.

Why this is a skill, not an agent

• Operator-invoked slash command. Reached as /nolte-media:gemini-image-handoff with a brief; the operator drives it directly.
• A human step sits between the two halves. The operator must paste the prompt into the UI and download the result before anything else can happen. An agent's fire-and-forget contract can't wait on a manual UI action.
• The deliverable flows back into the conversation. The copy-paste prompt block and the UI steps land in the operator's context to act on immediately.
• Counter-dimension: prompt authoring on its own could be an agent (graphic-prompt-generator is exactly that). The load-bearing dimension here isn't prompt quality but the interactive, operator-in-the-loop handoff — so this is a skill, and it delegates to graphic-prompt-generator when a brand-conformant prompt document is wanted.

German trigger phrases

• „generiere ein Bild über die Gemini-UI", „gib mir einen Prompt zum Einfügen in Gemini", „Gemini-Bild ohne Billing", „manueller Gemini-Weg, ich lade das Bild selbst runter"

How it works — the two halves

1. Author the Gemini prompt (automated)

Turn the brief into a prompt optimised for gemini-2.5-flash-image, following spec/design/gemini-image-generation/:

• Write it as narrative, descriptive prose (describe the scene; never a comma-separated tag list) and state the asset's intent or purpose.
• Front-load the subject, then action, location or context, composition, and style; be hyper-specific about material and texture; control the shot with photographic and lighting language.
• For in-image text, quote the exact words and name the font or style.
• Express any avoidance positively (a clean, uncluttered background over no clutter) — Gemini has no negative-prompt parameter.
• When the asset must be brand-conformant (uses the repo's brand tokens, will be published), don't hand-author it here: dispatch the graphic-prompt-generator agent with the target generator gemini-2.5-flash-image, then hand off the prompt block from the document it writes.

Present the result as a single fenced copy-paste block so the operator can grab it in one go.

2. Hand off to the Gemini UI (manual)

Guide the operator through:

1. Open the Gemini web UI — the Gemini app (gemini.google.com) or Google AI Studio (aistudio.google.com); a free Google account is enough, no billing.
2. Paste the prompt block from step 1 and send it.
3. Iterate in the chat if needed — conversational, one change per turn (keep everything the same, but make the lighting warmer); this is Gemini's recommended refinement path.
4. Download the image from the chat (the download / save-image control on the generated image).
5. Place and rename the downloaded file wherever the operator needs it — the skill does not move or record it.

Examples

• Read examples/01-hand-authored-prompt-handoff.md for the core one-off path: author a Gemini-optimised prose prompt from a brief, present the copy-paste block, and guide the UI download — no API, no sidecar.
• Read examples/02-brand-asset-delegates-graphic-prompt.md when the asset is brand-conformant and publishable, so the prompt is authored by graphic-prompt-generator and this skill only handles the UI handoff.
• Read examples/03-watermark-caveat-reroute.md when the asset must be watermark-free, so the unavoidable SynthID watermark makes the route wrong and the operator is steered to image-generate --provider cloudflare.

Hard rules

• Never call the Gemini API, request a GEMINI_API_KEY, or run the image-generate script for this flow; the whole point is the no-billing UI route.
• Never write the image or a <image>.meta.json sidecar — provenance and file placement are the operator's responsibility (this is the download-guidance-only ("Nur Download-Anleitung", the original requirement's wording) contract).
• Always state the SynthID watermark caveat: every Gemini UI output carries an invisible SynthID watermark. For a watermark-free commercial or blog asset, steer the operator to image-generate --provider cloudflare (FLUX.1-schnell, Apache-2.0, no watermark) instead.
• Always author the prompt to the Gemini baseline — don't paste a FLUX or SDXL tag-list prompt into the Gemini UI; a prompt isn't model-portable.

Gotchas

• Don't reach for image-generate --provider gemini here. That path makes a billed API call and writes a sidecar; this skill deliberately avoids both. They share the model but not the route.
• Brand assets go through graphic-prompt-generator first. It enforces the brand color contract and writes a durable prompt document; this skill then only handles the UI handoff for the prompt it produced.
• The watermark is unavoidable in the UI. There is no UI toggle to disable SynthID; if the asset can't carry a watermark, the route is wrong, not the prompt.
• Aspect ratio / resolution are the UI's to control. State the desired aspect ratio in the prompt; the chat UI doesn't expose the API's size parameters.