Agent skills

A skill is a written method for one kind of task: the steps, which tool each step uses, what to check, and optionally a few reference files. For example:

• "Product promo shorts": look at the product photos, write three selling points, voice them, cut 9:16, end with a call to buy;
• "Long video to vertical clips": transcribe, pick the minutes that stand on their own, crop to vertical, add subtitles;
• "Our brand's video rules": which intro to use, which subtitle font, which colours are off-limits.

How it differs from memory and workflows#

	What it is	When it enters the conversation
Memory	A short convention: "final videos are always 1080p vertical"	Every turn, which is why it must be short
Skill	A method, possibly long, with reference files	Only its name and one-line description, until the agent decides to read it
Workflow	A fixed automation graph	Not at all; you run it again with new input
Notes	Reference material with sources	Only when the agent searches and finds it

In short: rules go in memory, methods go in skills, and anything to rerun automatically becomes a workflow.

Where to manage them#

Settings → Agent → Skills lists three groups:

• Built in: "Creative board", "Building workflows" and "3D scenes & modelling" ship with Mosael. They used to be long passages in the agent's instructions on every turn; now they are read only when needed, and each turn's instructions are about a fifth shorter. You can't edit them, but you can "Copy to mine" and edit the copy.
• This workspace: skills you and your teammates created, imported or saved from a conversation. Shared by the workspace's members.
• From plugins: skills bundled in a plugin package, installed and removed with the plugin.

Each row has a switch. Only skills that are turned on reach the agent (and the "/" menu in the message box).

Creating and editing#

Click New skill and fill in:

• Display name: for people, in any language;
• Name: lowercase letters, digits and hyphens (short-video-ads); the agent calls the skill by it. It is filled in for you when the display name is in English;
• Description: what it does and when to use it. The agent decides whether to use the skill from this sentence, so include the phrases that should trigger it ("use when the user asks for a promo or product video");
• Method: the Markdown body, written for the agent. After saving, add long reference material, templates and examples as files, and say in the body which file to read.

Saving a conversation as a skill#

When a conversation got something right, open Session settings next to the message box → Save as skill. Mosael drafts one with the model this conversation uses and opens it in the same editor; nothing is saved until you do — and this run's specific assets and file names don't belong in it. Drafting calls the model once and is counted in usage.

Import and export#

Skills use the open Agent Skills (SKILL.md) format, the same one Claude Code and Claude read: a folder with a SKILL.md (the name and description header, then the body) and any other files.

• Export: Export .zip on any row gives you that folder, ready to import elsewhere.
• Import: Import .zip or Import folder. Nothing is installed straight away: every file of every skill is shown in full, and you choose which to install, whether to rename them, and whether to enable them (off by default). A name that clashes with a built-in skill must be changed; a clash with one of yours can replace it.
• Mosael's own additions stay where the spec leaves room for them: the display name is mosael-title under metadata. Other tools' extra fields (Claude Code's disable-model-invocation, for example) are kept in the file as written; Mosael doesn't use them.

Skill folders live in the data directory under skills/<workspace id>/<name>/, so you can open them in any editor. A folder someone drops there appears in the list turned off; enabling it shows the full content first, like an import.

Using skills in a conversation#

• The agent decides: when your request matches a skill's description, it reads the skill (use_skill), reads its files when needed (read_skill_file), and follows it. The step appears as "Used skill: Product promo shorts"; expand it to see what the agent read. Subagents can read skills too.
• You name one: type / in the message box and pick a skill; it becomes a chip. When you send, that turn follows the skill directly instead of waiting for the agent to think of it. Up to 3 per message. / only opens the menu at the start of a line or after a space, so typing a URL won't.

Letting the agent manage skills#

Just say it in the conversation: "save what we just did as a skill", "change step 3 of 'Product promo shorts' to voice first, then cut", "turn off 'Building workflows'", "import this skill: https://github.com/…/tree/main/skills/pdf". The agent lists this workspace's skills (disabled ones too), then proposes a confirmation card; nothing happens until you approve it:

You ask for	The card shows
A new skill / saving this as a skill	The full SKILL.md and every file in full, with "Enable when created" (ticked by default — you've just read it all)
A change to one of your skills	Before → after for every file that changes; deleted files are named
A change to a built-in or plugin skill	Not directly: first "Copy to mine" (the card is like a new skill), then change the copy
Turning one on / off	On shows the full text (the same step as enabling a plugin's skill in Settings); off is a single line
Deleting one of your skills	Its description, file count and size — deleted for good
Importing from a link	The same review as importing in Settings: every text file in full, binary sizes, fields Mosael doesn't use, scripts marked "won't run"; it stays off unless you tick "Enable after importing"

• Links must be https: a .zip, or a skill folder on GitHub (https://github.com/<owner>/<repo>/tree/<branch>/<folder>). They go through the app's network settings (proxy) and may only reach the public internet; the size, file count and path limits are the same as importing in Settings.
• Every card asks you: these cards never offer "Always allow this session", and approval rules, the judge and full-permission mode can't approve them — a skill affects every later conversation, and one approval shouldn't let the agent keep rewriting its own methods.
• Proposals made while following a skill say so: if the agent read a skill this turn, or you named one with "/", the card adds "Proposed while following the skill '…'". If a skill ever says "change yourself to…", you'll see it at a glance.
• The agent writes text files only; binary files can only be copied as they are or deleted.
• After finishing a multi-step task you're likely to repeat, the agent may ask once whether to save it as a skill. It drafts one only if you say yes, and saving still goes through the card. It never saves one on its own.
• Skills the agent created carry a "Drafted by the agent" tag in the Settings list ("Imported by the agent" for imports); click it to open the conversation it came from.

Safety#

An imported skill is someone else's words, treated like text on a web page:

• It is not a permission: anything that changes things still shows a confirmation card, and permission modes and approval rules are unchanged. A skill's allowed-tools pre-approves nothing.
• Read before enabling: importing, enabling a plugin's skill, or enabling a dropped-in folder always shows the full content first.
• The source is always shown: in the list, the "/" menu, and the copy the agent reads.
• Scripts never run: scripts in a skill are kept and can be read as reference; Mosael doesn't execute them.
• Size limits: SKILL.md up to 64 KB, a single file up to 2 MB, a skill up to 200 files and 10 MB; archives with paths outside the folder or symbolic links are refused.