| name | curation-digest |
| user-invocable | true |
| argument-hint | [{"optional":"style or artist to focus"}] |
| description | Browse the TITLES (titles.xyz) feed, judge the best work in a style or by an artist, and save it to a collection with a short taste report. Use when the user says: "find the best X on TITLES", "curate a moodboard", "save some references", "what's good in [style]", "build me a collection", "show me strong work by [artist]". Searches the published feed, looks at the candidates, saves the keepers, and summarizes what makes them work. Mostly read-only — no generation cost. Runs on the TITLES MCP — if TITLES tools are missing, connect mcp.titles.xyz/mcp first (see titles-setup). NOT for: generating new images (generate-image / promo-pack), or exploring the library by generating (style-explorer).
|
curation-digest
Find the best work on the TITLES feed in a style, save it, and say why it's good. Near-zero cost — this is browsing and saving, not generating.
Get connected
Check the tool list for TITLES tools (names contain titles_). If missing, hand off to the titles-setup skill and stop.
1. Frame the search
Get the target — a style ("analog collage", "low-poly retro"), an artist, or "what's good right now".
Size the keep-count to the pool, don't fix it up front. Curation is the cut: keep roughly half of what you look at, floor of 3. If a pool of 11 yields 8 saves you collected, you didn't curate. Tell the user the ratio in the report ("5 of 11").
2. Build the candidate pool
titles_search_feed does not take a free-text query — it filters by model_id, username, media_type, and remix flags, and sorts by feature or date_created. So route by target:
- By style → first
titles_search_models with the style terms to get the artist models in that style, then titles_search_feed({ model_id }) for each of the top matches. This is the main path; a style has no single feed query, it's a set of models.
- By artist →
titles_search_feed({ username }).
- "What's good" / a moodboard with no fixed style →
titles_search_feed({ sort_by: "feature" }) for the featured pool.
Keep page_size at 10 or less and page with the cursor. Feed items are heavy (~4KB each — full asset metadata, avatars, embedded tool icons); page_size: 20 on a prolific artist overflows the tool-result limit and dumps to a file. media_type: "image" (or "video") narrows it.
3. Look at the work, then cut
The feed tools return no images — titles_search_feed and titles_get_feed_item give metadata only (unlike your own generations, which inline previews). To actually see a candidate, fetch its thumbnail from the public CDN and view it:
curl -sS -o cand.webp "<output.asset.thumbnail_url>" # public, no auth
sips -s format jpeg -Z 420 cand.webp --out cand.jpg # macOS; any converter works
Then read the file. Do this for every candidate before cutting — judging from titles and descriptions alone isn't curation.
Pick on craft, coherence with the brief, and variety (don't keep six near-duplicates). Common disqualifiers worth watching for: garbled hallucinated text where a model tried to render type, and compositions with unfilled gaps or unfinished edges. titles_get_feed_item adds the original prompt and an is_saved flag when you want more context on a finalist.
Credit runs two layers deep on TITLES, and both belong in the report: the publisher (owner.username) and the model's artist (samples[].name where model_type is artist) — often different people. When a piece is a remix, the source work's creator is a third credit, usually named in the description.
4. Save the keepers
Save each keeper to the user's library with titles_set_collection({ kind: "saved", target_id: <published_id>, state: true }) — target_id is the feed item's published_id (the id on the search result), not the output_id. TITLES has one saved library, not named per-brief collections, so the grouping is the report you write in step 5, not a container in the product. The response's changed tells you whether it was newly saved or already there.
5. Deliver the taste report
A short write-up is the collection: the keep ratio ("5 of 11"), then each pick with both credits, its published_url, and one line on why it earned the spot — pointing at what's actually in the image, not the title. Say which ones were newly saved vs. already in the library (the changed flag). Name what you cut and why; the cuts prove the picks. Close with the pattern you noticed across the set — that's the part a list can't give them. Then offer the next step: generate in this direction (promo-pack), explore the range (style-explorer), or follow an artist to keep their work surfacing (titles_set_following).
Etiquette
No generation, no spend — so no cost gate. Be opinionated in the picks; a curation report that saves everything isn't curation.
If the user pivots from curating to making ("now generate one like these"), that's promo-pack (make a set) or style-explorer (roam styles) — hand off rather than generating here, so the cost gate those skills carry isn't skipped.
