The eight Versely skills and when each fires

The eight Versely skills are not eight synonyms for "generate." Each pack is a procedure for a different pipeline shape, and the agent chooses among them from the wording of your request. Ask for "a song" and you may land in generate. Ask for "extend this track and split the stems" and you want music. Ask for "a five-scene film that chains each shot from the last frame" and you want movie. The skill that fires is the one whose description you matched, so phrasing is the whole interface.

Install once with npx skills add AI-XLabs-Innovation/versely-skills, or as part of npx @versely/cli setup. The skills page is the product surface. Pair them with the MCP server so the agent has callables as well as instructions. Then talk in jobs, not in endpoint names.

How an agent picks a skill

Each pack is a folder with a SKILL.md. The frontmatter description is the matcher: a short paragraph that names the work and the "use when" clause. Agents that support the Anthropic skills format load those descriptions and surface the pack whose clause fits the turn.

That means the agent is not reading your mind. It is matching vocabulary.

• "Generate," "create," "produce," plus an image, video, or sound effect: generate.
• "Slideshow," "carousel," "reel" as a stack of stills: slideshow.
• "Multi-scene," "short film," "storyboard," "chain the frames": movie.
• "Talking head overlay," "UGC," "burn captions": ugc.
• "Extend the track," "split vocals," "Suno," "stems": music.
• "Post this," "schedule to Instagram," "publish": social.
• "How did that post do," "what's trending": analytics.
• "Do the whole pipeline," "a week of content," "generate then caption then post": content-pipeline.

If two could match, the more specific pack should win, but only if your wording is more specific. "Make me a video" is generate. "Make me a three-scene video where scene two starts from scene one's last frame" is movie. The extra constraints are what move the matcher.

The eight, and the request that fires each

Skill	Fires when you are asking for	Does not fire for
versely-generate	One-shot images, video clips, lipsync, TTS, sound effects, a simple Suno track	Multi-scene chained films, slideshows-as-objects, stem splits, publishing
versely-slideshow	A stack of stills, a carousel, a stills-to-video reel	A single generated image you will post as-is
versely-movie	A project with scenes, retries per scene, previous-frame chaining	One clip from a prompt
versely-ugc	Overlay a talking head, burn captions, auto-time speech, compose clips under an overlay	Generating the base product video (that is generate)
versely-music	The full Suno surface: extend, mashup, replace a section, stems, MIDI, personas, covers	A one-line "make a lo-fi bed" that generate can already do
versely-social	Publish, schedule, list, or delete posts on connected accounts	Generating the asset you are about to post
versely-analytics	Metrics on Versely-published posts, a trend analysis, a public trending URL	Publishing, or generating a follow-up clip
versely-content-pipeline	A multi-step flow that chains the packs above, with a budget step first	A single generate, which should stay in generate

Concrete prompts that land in the right pack:

Generate. "Generate a 9:16 clip of a ceramic pour-over on a walnut counter, morning window light, five seconds." Generate a video from text is the same job in the in-app agent.

Slideshow. "Make an 8-slide reel about celebrity-adjacent productivity myths, 9:16, captions burned into each still, then convert it to video."

Movie. "Create a three-scene 16:9 film. Scene one is a traveler on a misty path. Scene two chains from that last frame onto a stone bridge. Scene three ends on a castle." The movie pack is explicit that you poll GET /api/v1/movie/:id/status, not /status/:requestId per scene. Create a multi-scene AI movie is the in-app version of that project.

UGC. "Overlay my talking-head clip in the bottom-right of this product video, small, then auto-caption the speech in yellow." Generating the product video first is generate; the overlay is ugc.

Music. "Extend this Suno track from 90 seconds, then separate vocals." A first generate of an instrumental bed can stay in generate. Anything past /suno/generate belongs here.

Social. "Post this URL to the connected Instagram and TikTok accounts with this caption. If nothing is connected, stop and tell me." The pack lists accounts before it posts. Post or schedule is the in-app twin.

Analytics. "Pull engagement on last week's Versely-published posts, then analyse 'AI tools for solo founders' on TikTok and YouTube."

Content-pipeline. "Turn this product brief into a five-image carousel, convert it to a reel, caption it, and schedule it to Instagram and TikTok tomorrow 9:00 UTC. Check credits before you start." That last sentence is doing real work. The pipeline pack opens with a budget step because a multi-step flow that dies at step six has already spent steps one through five. Set up a full content pipeline is the in-app form of that request.

Phrase the job, not the tool

The matcher is vocabulary, so help it.

Name the shape. "Carousel," "chained scenes," "talking-head overlay," "stems," "schedule this post," "trend report." Those words sit in the descriptions. "Make me content" sits in none of them and will drift toward generate.

Name the next step only if you want it in this turn. "Generate a clip" stays in generate. "Generate a clip and post it" is a pipeline. If you did not want publishing yet, do not mention publishing.

Do not name the endpoint. "Call /api/v1/movie/create" is how you skip the skill and get a raw curl. The pack exists so you do not have to. If you want movie semantics (per-scene retry, auto-combine, chained frames), describe those semantics.

Split generate from music on purpose. The generate pack defers anything beyond /suno/generate to the music pack. "Make a 30-second instrumental" can stay. "Cover this upload in acoustic guitar, then give me MIDI" cannot.

Split generate from movie on purpose. One prompt, one clip: generate. A project you will retry per scene: movie. If you poll /status/:requestId on a movie you will be looking at the wrong object.

Scopes follow the same split, which is why a key that only has generate will 403 on slideshow and social even if the skill fired correctly. Generate covers /generate/*, /suno/*, /movie/*, /ltx-batch/*. Slideshow is slideshow. UGC is ugc. Social is post plus manage_accounts. Analytics is analytics. Pipeline needs the union of whatever steps it will run. The CLI's setup mints a key; it does not magically grant every scope. Check what you created.

What the pipeline skill is for

versely-content-pipeline does not own a unique API. It orchestrates the others: generate then overlay then caption then post; slideshow then convert then schedule; a week's calendar of images with scheduled_at timestamps. Its job is the budget, the order, and the recovery rule.

The recovery rule is the part people skip. If step three fails, do not re-run step one. Intermediate URLs stay valid. Retry the failed step with the same input. Report partial results. A generated image that never posted is still an image.

Use pipeline when the turn actually spans packs. Use generate when it does not. Pulling pipeline for a single Flux image adds a credit speech and a social-accounts lookup you did not ask for.

If the agent picks wrong, the cheapest correction is a follow-up that names the shape: "Not a single clip. A three-scene movie that chains frames." You should not have to say the skill's filename. You should have to say the job.

FAQ

Do I need all eight installed?

The install is all eight. You do not have to use all eight. An agent with no connected social accounts can still load versely-social; the pack's own pre-check is "list accounts, and if there are none, stop." Unused packs are idle instructions, not idle spend. Spend starts when a generate (or post, or analysis) actually runs, and every one of those costs credits.

What if the agent uses generate for a movie?

Stop it and restate the constraints: multiple scenes, per-scene status, chain from the previous last frame. Generate will happily fire N independent /generate/video calls and give you N clips with no project object and no combine step. That is the symptom. The movie pack exists because /movie/* is the first-class project API for that shape.

Can I invoke a skill by name?

In agents that surface /versely-* slash commands after install, yes. The reliable method across agents is still the description match. Write the job so only one pack's "use when" clause fits. Slash commands are a shortcut, not a requirement.

Skills or the in-app agent?

Same catalog, same balance. Skills run inside Claude Code, Cursor, Codex, and the other agents the CLI detects. The in-app agent already has routing of its own and does not load these files. If you are in a coding agent, install the packs. If you are in the product, just ask.