The catalog

How you’ll use it

Image: thirteen

Skill	For
Product Photography	The product as the subject: packshots, lifestyle, hero shots.
Social Graphics	Still imagery for feeds and stories.
Web Graphics	Page furniture: heroes, section imagery, cards.
Email Graphics	Imagery that survives an email client: compression, dark mode, images-off.
Display / Retargeting Ad	Paid banner placements at fixed IAB sizes, in a sub-second viewing budget.
Print Collateral	Anything going to press: trim, fold, gamut, rich black.
Billboard / OOH	Read at distance, in daylight, in seconds.
Editorial / Blog Imagery	Illustrating a thesis rather than a topic.
Avatar Portrait	A person’s likeness, held stable and usable downstream.
Character Reference Sheet	One character across multiple views, identity held between them.
Reference View	One view of one state of a subject. An input, not a deliverable.
Location Plate	One plate of one location, anchored to something unmistakable. Also an input.
Location / Environment / Set Design	The place itself, built to be shot in.

Video: five

Skill	For
Social Video	Short-form, fast open, chrome-aware.
Video Frames	Anchor frames for a move: the start and end a model interpolates between.
Cinematic / Brand Video	One considered shot, with real camera vocabulary.
Multi-Shot Video	Several shots that have to agree with each other.
Avatar / Talking-Head Video	A person delivering a script to camera.

Audio: three modes, one Skill

Audio has three modes (Music, Speech and SFX) and one Skill, which belongs to SFX. Music and Speech have no Skill of their own yet: the agent composes for them from your brief and the request alone, the same way any mode behaves before a Skill is confirmed.

Skill	Mode	For
Physical Event	SFX	A sound caused by something happening: an impact, a mechanism, a material.

Important things to keep in mind

Three of the image Skills make inputs, not outputs. Reference View, Location Plate and Character Reference Sheet exist to produce the assets a shot list asks for (one view of one state, one plate of one location) so that later shots have something consistent to stand on. Their failure modes are about that: inherited pose, environment leak, a plate mistaken for the shot. If you are judging one of them as a finished picture, you are judging it against the wrong standard.

Edit Image and Edit Video are not in this list, and that is deliberate. They are Operations (you send an existing image or video back for a new generation with a Key Prompt), not Skills describing a kind of work. The distinction matters because a Skill answers what am I making, and an Operation answers what am I doing to something I already have.

The count is nineteen, and the catalog in the code currently carries twenty-one. The two extra entries are Edit Image and Edit Video, which were ruled Operations on 2026-09-03 and have not yet been removed from the generated catalog. This page documents the nineteen.

The catalog is a floor. These are the kinds of work worth having a considered answer to out of the box. Everything about writing your own applies equally to forking one of these and changing what does not fit your house.

Editing is an Operation, not a Skill

A Skill is a kind of work: a product shot, a social video, a character sheet. Editing is not a kind of work. It is something you do to a product shot, or a social video, or a character sheet. Filing it beside them would mean every Skill needed an editing twin.

The conversational path is the normal one. Attach an existing image or video and say what to change: “make this warmer”, “change the label”. Instruction language plus an attached asset is signal enough; nothing needs selecting.

If your attachment is ambiguous, you get one question. An asset attached with no instruction, or an unclear one, produces a single clarifying question: use this as a reference, or edit it directly? Those are genuinely different requests: a reference informs a new generation, an edit changes the thing you gave it.

The Operation select pre-resolves that. It sits beside the Skill picker in the agent dialog, so you can say up front that this is an edit and skip the question. That is the whole reason the control exists: discoverability without needing to already know the conversational pattern.

These used to be Skills, and the change is recent. Edit Image and Edit Video appeared in the Skill list purely so the capability was discoverable. EDIT OPERATION 1.0 (2026-09-03) moved them out: discoverability now sits on the Operation select, “from a control that says what it does”, and the Skill list went back to describing kinds of work.

Edit Image is on the Operation select. Edit Video currently appears nowhere. That asymmetry is real and is not a documentation gap. No live model has a video-edit endpoint, so video editing is an open gap, as How we track models says.

The generated catalog still carries both as Skills. catalog.generated.ts holds edit-image and edit-video with routing and failure modes, which is why a count of the file returns twenty-one against a user-facing nineteen. Recorded, not yet cleaned up. It matters beyond arithmetic: those entries carry routing, so a request can in principle route to an Operation as though it were a kind of work.

No drawn-mask editor, deliberately. Leading edit models increasingly handle plain-language edits without a mask, and the ambiguous cases a mask solves are a shrinking target. Clarifying questions cover most of the same ground with no new interface.