| name | promo-pack |
| user-invocable | true |
| argument-hint | ["campaign goal"] |
| description | Generate a set of promotional campaign assets on TITLES (titles.xyz) in one artist's style — cover, social posts, banners, story crops — each sized for wherever the campaign runs. Use when the user wants marketing or promo visuals as a set: "promo pack", "campaign assets", "album promo", "launch visuals", "social media pack", "assets for my release", "marketing set", "cover plus socials". Asks the goal, then the style; explores a few artist models within a budget; iterates the direction with you; then produces the full set at the right aspect ratios, and can add text, edits, or a final upscale. Runs on the TITLES MCP — if TITLES tools are missing, connect mcp.titles.xyz/mcp first (see titles-setup). NOT for: a single image (generate-image), one edit to one image (edit-image), or a video set (motion-pack).
|
promo-pack
One goal → a coordinated set of campaign assets in one artist's style, at the aspect ratios the campaign needs. Explore cheap, lock a direction, then produce the set.
Get connected
Check the tool list for TITLES tools (names contain titles_). If missing, hand off to the titles-setup skill (connect https://mcp.titles.xyz/mcp, OAuth, free to start) and stop. Don't fall back to non-TITLES tools.
1. Goal first, then style — ask, never invent
Two asks, in order, and both are the user's to give — if either is missing, ask and stop. Don't infer a goal or a style from earlier conversation, and don't treat a category name as a brief.
- The goal / end product — what the campaign is for and where it runs (a record release, a product launch, an event). This decides the asset list.
- The stylistic direction — an artist, a reference, or a mood, in the user's words.
From the goal, decide the asset list and aspect ratios, and say them back. This is yours to judge from where the assets will live:
- A record promo typically wants 1:1 (cover), 4:5 (Instagram post), 16:9 (YouTube/site banner), 9:16 (story/reel).
- Other campaigns need different ratios — pick what fits the channels the user named. Don't force the record set onto a billboard or a web hero.
2. Find models + write the exploration prompt
titles_search_models with style + subject nouns from the goal and style (never guess model names — the catalog changes). Gather several fitting artist models, each with its agent_description / tags, artist name, and model_url.
Write one campaign prompt — the hero concept — that you'll run across all the models you explore, so the comparison is about the artist's style, not the prompt.
3. Budget → how many models to explore
Ask the user's exploration budget. Then probe one sample (step 4's first call) to get the real per-sample cost, and compute how many models fit: more budget = more models, but cap around 5–6 (beyond that the comparison is noise) and floor at 2. Tell them: "at ~$X/sample, $BUDGET explores N models."
The budget covers exploration only. The final set and any edits are quoted and confirmed separately, later. Say so now so the number isn't mistaken for the whole job.
4. Explore — one sample per model
Generate one sample in each of the N models: the same exploration prompt, the same ratio (use the primary deliverable's ratio — usually 1:1 — so each sample doubles as a real cover candidate). Run one at a time on a shared session_id; the awaits afterward can run in parallel.
Read cost_usd off each submit response to track spend (older servers nest it as cost_quote.total_usd — read whichever comes back). Present the samples labeled by artist, each credited, via the canvas link and downloaded files (see Delivery).
5. Feedback rounds
Ask for feedback and iterate — adjust the prompt, swap a model, try more of a direction — for as many rounds as the user wants, until they're happy with one direction (a single model + prompt). Every round spends, so state the running total each time and flag before a round would pass the exploration budget.
6. Produce the full set
Direction locked. Now quote the set separately: project the chosen model's per-image cost across the asset list, per aspect ratio (cost scales with pixel count — a 1:1 runs richer than a 16:9/9:16). Show the total and get a go — this is fresh spend beyond exploration.
Generate each asset at its ratio in the locked model, one subject per role — hold the style constant (the model does that for free) and give each asset its own subject inside the same world, so the set is different images that belong together, not one image resized. One at a time, shared session_id, running cost reported. Deliver via canvas + downloads, and offer a feedback round on the set too.
7. Text & edits — across the whole set
When the set is right, ask if they want to add text (title, tagline) or make edits. An edit like adding a title applies to every image the user accepted, not just one — carry it across the whole set unless the user names a subset. Ask which tier:
- Higher quality (pricier) → Nano Banana Pro. Resolve it live:
titles_search_models({ query: "nano banana", operator: "imgEditNode" }). Best for text and precise edits — the right call for adding a title.
- Budget → Seedream 5.0 Lite. Resolve live:
titles_search_models({ query: "seedream lite", operator: "imgEditNode" }).
Quote the whole step before running: the per-image edit cost × the number of accepted images being edited, shown as one total — this is fresh spend across the set, not a single edit. Get a go.
Then run the edit on each accepted image as its own base: titles_edit_image({ output_id: <that image>, prompt: <only the change they asked for>, model_id: <resolved id>, session_id }), one at a time on the shared session_id. Keep the change identical in content but adapt its placement per asset — a tagline sits differently on a 1:1 cover than a 9:16 story. Report the running total across the batch; iterate as asked.
8. Publish (offer)
Offer to publish selected finals to the TITLES feed with artist credit: titles_publish, showing the name + description first and getting the user's go. Offer it — don't assume — many promo assets are for off-platform use.
9. Upscale (offer) — final step
As the last step, offer to upscale the finished assets to higher resolution for print and delivery. Use SeedVR — never Bria: resolve it live with titles_search_models({ query: "seedvr", operator: "imgUpscaleNode" }) and pass that model_id (the Bria creative upscaler is a different, pricier model — don't use it here).
Upscaling applies to all the accepted/final images (or the subset the user names). Quote the whole step first — the per-image upscale cost × the number of images, as one total — and get a go. Then titles_upscale_image({ output_id: <each final>, model_id: <SeedVR id>, session_id }), one at a time on the shared session_id, running total reported. Deliver the upscaled files via Delivery below.
Delivery — canvas + files
- Always give the
session_url (the canvas). Raw output/CDN URLs 403 outside the browser.
- For the files,
titles_download_asset({ output_id, format }) returns a short-lived (~5 min), no-auth download URL:
- Shell host (Claude Code / Codex):
curl each to a named folder on disk and tell the user the path.
- Chat host (claude.ai / mobile — no filesystem): hand them the link(s) to click before they expire; re-fetch if one lapses (cheap).
Cost & etiquette
Every generation returns cost_usd on submit — keep a running total and remind the user at each gate and each feedback round. Within an approved budget, fire fast; don't re-ask per generation. One canvas link, the artist credited, the set laid out — no play-by-play between generations.
