Skip to main content
A skill is a reusable creative recipe — a named markdown playbook the agent follows end-to-end when you invoke it. Think of it as a prompt template with teeth: structured enough that the agent executes it reliably, flexible enough that you can still hand it new subjects and contexts. Good candidates for skills:
  • A “hero shot” recipe that always produces the same composition, lighting, and finish
  • A “story reel” that always builds a 4-scene narrative from a single product
  • A “platform pack” that turns one idea into Instagram square + Reel + LinkedIn variants
  • A “brand audit” that runs a consistent visual analysis across saved inspirations
If you’re copy-pasting the same multi-step prompt more than twice, it belongs in a skill. Skills apply at the stage they describe. A recipe for inspecting a finished render should not be forced onto an early draft. Memo compares the task with each skill’s stated purpose, loads the matching recipe, and keeps it through relevant follow-ups. Large libraries are searched through short descriptions; the full text of every skill is not added to every prompt.

​
Where skills live

Open Brand Kits → Skills for general skills and Memo starters that work across brands. Brand-specific skills live in their Brand Kit → Guidance & skills. Turn on Include brand skills in the main library to browse them alongside general skills; each shows its kit name and a button to open that kit. These are two views of the same saved skill: edits from either place update one record. Search and the sidebar count follow the selected scope. Collections contain Images and Rules; they do not own or recommend skills. Two ways to view the library, switchable with the Grid / List toggle next to search:
  • Grid — compact entries arranged in two columns at wide desktop widths, showing a thumbnail, name and use description. A skill with a fixed image shows that image; every other skill gets a Memo gradient tile. Relevant brand and draft labels stay visible.
  • List — the same compact entries in one column, with a row menu (⋯) for Open and Delete.
Your layout choice is remembered per browser. Skills created by a connected agent remain attributed to whoever connected it in the saved record. From here you can:
  • Create with the New skill entry at the top of the library, the same ”+” entry Collections uses.
  • Read a skill, then choose Edit to change its instructions or references.
  • Search by name or content as the library grows.
  • Delete skills you do not use, after confirmation.
When the library is empty, the Create Skill action appears directly in the first-run state so there is still one obvious way forward.

​
Memo-official starters

Every workspace ships with a starter pack authored by Memo. They sit in the same library, are read-only (open one and Duplicate it to make your own version), and connected agents get them through list_skills / get_skill like any other skill:
  • Naming Convention — the tag grammar and never-overwrite rules that keep a multi-asset production findable.
  • Scene Engine — stress-test a scene or sequence against Goal, Obstacle, Tactic, Reversal, Value Shift before spending credits.
  • Acting Task — direct character performance with playable tasks instead of emotion labels.
  • Blocking Map — exact multi-character staging via a color-coded schematic sent as a reference.
  • CINEDANCE — production-ready cinematic video prompts for one shot: references, blocking, optics, physics, lighting.
  • Timed Prompts — timecoded multi-shot video prompts for Seedance 2.5 and similar models: several shots with exact timecodes, fixed dialogue, match cuts, continuity across locations, blocking maps. Elements are named by their Cinema Kit handles; the result is a brief, one scene per part. Its failure fixes, blocking method and reframe recipe are separate notes the agent opens when needed.
  • Motion Design — render-ready 15-second launch and motion-graphics video prompts: one core mechanic, locked palette, beat sheet, copy-list contract, fully silent.

​
The editor

Open a skill to read its description and instructions. Choose Edit to make changes. Memo starters and skills opened by viewers remain read-only. Choose New skill to create one. The main form has three fields:
  • Name — a short title you can recognize and select in chat.
  • Use this skill for — one sentence describing when it is useful, such as “summer menu photos for social media.” Agents see this description when discovering skills; it is not an automatic trigger.
  • Instructions — what to do and what the result should be. Plain English is fine. Type @ to reference brand assets, colors or instructions.
Add a fixed image is optional. Use it only when this recipe should use the same image every time, such as an approved product or character. For other images, choose references when working on a task. Brand assets attach directly; picks from Generations or Inspirations are first saved as brand assets. The skill belongs to the place where you create it: a Brand Kit or the general workspace library. Editing an existing skill keeps its location. There is no separate location chooser or expected-result field; describe the finished result in Instructions. Older saved expected results appear there too and move into the document when you save. Existing images, model pins, notes and inactive legacy collection associations are preserved. Put requirements that always apply in rules and the steps for a particular task in the skill. Choose the model where you generate; existing saved model pins are preserved.

​
Saving and long instructions

Save skill makes a new skill available after checking its use description, nonempty instructions, reference structure and size limits. Save draft keeps unfinished work out of the agent’s runnable library. Older skills retain their existing compatibility when saved. There is no minimum word or step count. Available new skills allow 12,000 characters of instructions; drafts allow 50,000. The use description allows 500 characters. Text is never silently cut to fit. If instructions exceed 12,000 characters, a small help panel appears. Separate examples and notes proposes moving sections headed Examples, References, Background or Notes into attached notes, keeping their exact text. Review the proposed instructions and notes, then choose Apply separation or Keep original. Keep requirements the agent must always follow in the main instructions. If no sections can be identified, save a draft while shortening the text. No model runs and no credits are spent. Examples and notes appears only if the skill already has notes or you have separated them from long instructions. Expand it to read or edit them. The agent can fetch these notes when needed; they are not all loaded into every task. Notes allow 12 entries, 20,000 characters each, 100,000 total. Failed saves keep your edits. If another editor changed the skill, reconcile with the latest version before saving again.

​
Ask the agent instead

In Studio chat, select an image on your canvas and ask the agent to turn it into a skill (“save this as a reusable recipe”). Save Skill puts the proposed recipe in your library as a draft. Open it to review its use description, instructions and references, then choose Save skill.

​
Invoking a skill

Two ways to call a skill in Studio chat:

​
Slash command

Type / in the composer. A picker appears showing available skills (and Workflows when enabled), filtered as you type. Drafts are excluded and cannot be invoked through a stale picker selection.
The agent reads the playbook as the brief for the turn — your free-text addition is the subject. Referenced images are injected automatically: the agent sees them, and their IDs are attached to every generation server-side, so brand consistency doesn’t depend on the model remembering.

​
@mention

Type @ followed by a skill name — useful when combining a skill with a brand kit mention in one message (@Acme Brand @story-reel ...).

​
Plan-then-execute

For skills that produce multiple deliverables (a storyboard recipe, a platform pack), the agent proposes a numbered plan first — usually as a creative brief on your canvas. You review, adjust ratios and references, and click Generate. The skill’s prose controls the plan; your confirmation controls execution. Skills slot neatly into the Creative Brief flow.

​
How aspect ratio and quality get chosen

Skills hold the creative intent — style, references, directions. They don’t hold a fixed aspect ratio or quality, so one /hero-shot can produce a 1:1 feed post AND a 16:9 banner:
  • Single-output skills — the agent asks once for ratio and quality, then generates. If the playbook already specifies a ratio (“produce a vertical 9:16”), it skips the question.
  • Multi-deliverable skills — the agent builds a brief with one deliverable per output, each with its own ratio and quality row, enforced server-side at generation.

​
Combining skills with brand kits

A General skill works with any brand: @mention a kit in the same message and the skill provides the recipe while the kit provides the identity.
A brand skill belongs to that Brand Kit and is labeled accordingly. Keep reusable task instructions here, and put collection-specific requirements in the collection’s Rules. For example, a product-photography skill explains how to compose a shot; a device collection rule says where its logo must appear.

​
How the agent chooses

You can choose a skill yourself, or let the agent find one for your request. Both paths load the same saved instructions, attached images, linked instructions, colors and model settings. The agent first sees short descriptions from your workspace, the current Brand Kit and Memo starters. It searches names and use descriptions, then reads only relevant recipes. Up to 24 matches are shown at a time; this is a page size, not a limit on your library. Extra notes stay separate until needed. Choosing no skill is also valid. When the agent chooses, chat shows Using [skill name]. If two approaches would produce meaningfully different results, it should explain the difference and ask through the normal question card. It waits for your answer before selecting or generating. You do not need to approve routine skill browsing. Brief and credit approvals still work as before. Selecting a collection changes the applicable rules, not the skill pool. Changing brands clears an inherited skill choice. Follow-ups keep the appropriate skill based on the task, not message length. Long refinements can continue the same task; short unrelated requests start another. The saved task survives reloading the chat, while skill instructions and dependencies are reloaded from their current saved source. An approved brief keeps its selected references and settings: start a new brief to change its skill. If a skill has a missing image or linked instruction, Memo asks for it to be repaired instead of silently leaving it out. A skill can prepare up to 16 attached image IDs and five linked instructions, with up to 12,000 characters of linked instruction text in total. Individual generation models can accept fewer images. Move optional detail into extra notes.

​
Skills over MCP

Connected agents (Claude, or anything speaking MCP) work with the same library:
  • list_skills searches names and use descriptions and returns paginated summaries. Set a Brand Kit to include its skills plus general skills; omit the Brand Kit or use null for general skills only. Discover Brand Kits first when looking for a brand-specific workflow. Drafts are excluded.
  • get_skill returns the playbook, a supporting-reference catalog and prepared attachments/settings for a ready saved skill. Reading it does not select or run it. Set brand_kit_id to check that it fits the current brand. Use include_draft for editing reads.
  • get_skill_resource reads one supporting reference in bounded chunks, with an explicit continuation when needed.
  • create_skill / update_skill let an agent capture a workflow you refined in conversation, or improve an existing playbook — attributed to you in the saved record. Updates keep an automatic copy of the previous version, so an agent rewrite is never a one-way door.
After reading and choosing a saved skill, the connected agent passes its skill_id to generate_image or generate_video. Memo adds its saved image references and model pin before calculating the estimate. A model explicitly chosen for that generation overrides the saved pin. The agent still follows the skill’s prose to compose the request; selecting a skill does not turn all its steps into an automatic job. The estimate names the skill. Confirmation checks that the prepared recipe and dependencies still match; if they changed, the agent requests a fresh estimate. Nothing is submitted by reading or selecting a skill. The external chat app controls how it displays questions and approval; Memo supplies the tool data. New MCP-created skills default to draft. Agents can explicitly request ready status with the required use description and bounded instructions. Updates to existing skills preserve omitted configuration fields, including inactive legacy collection associations and supporting references. New collection associations are rejected.

​
Editing skills

Changes apply to future invocations immediately — already-generated work doesn’t change. If a skill produces inconsistent output, tighten its prose: be specific about deliverable count, style anchors, and required references. Open the skill, then choose Edit. Images and extra notes stay attached when you save. Copy link next to Edit gives you a link to the skill to paste to an agent or a teammate.

​
Tips

  • Start from a prompt you’ve already perfected. If a multi-step prompt keeps working, paste it into a skill and name it.
  • Make skill names short. You’ll type them after a slash. /hero-shot beats /product-hero-shot-three-angles.
  • Keep each skill single-purpose. One recipe per skill. If it’s drifting into “and also does X,” split it.
  • Write the playbook in the agent’s voice. Be specific about what to produce; leave room in how, so your brand kit can do its job.
  • Attach the concrete references. If the prose mentions “the character” or “the product,” make sure that image is actually attached — otherwise the agent has prose pointing at nothing.
  • Separate requirements from procedure. Store brand-wide requirements in the Brand Kit’s rules. Keep steps for this particular recipe in the skill, and detailed examples in supporting references.
  • Test each skill with two different subjects before relying on it. If output quality differs wildly between subjects, the prose needs tightening.