Skip to main content
Agents are specialized environments with their own instructions, knowledge, integrations, and model. Skills are reusable instructions that load when they match the request, such as tone of voice or output format. You can attach Skills to an Agent so both work together. If you move generic rules like brand guidelines into Skills, Agents stay focused on what makes them specific. If your team opens an Agent only to get properly formatted output, that should be a Skill instead. If people have to go somewhere for something that should always just work, it belongs in a Skill.
Use a Skill for rules that should apply in regular chats, such as presentation branding. Use an Agent when you need a full process with its own knowledge and tools, and attach Skills when those rules should apply there too. Start with one focused Agent and the matching Skills rather than many Agents.
Project instructions apply only to project chats. Skills apply to normal chats and are not limited to a single project.
Skills are model-agnostic.
In a regular chat, ask the built-in Skill Creator to create a Skill, including from the current conversation. Chat shows Save this skill? and writes nothing until you click Save skill. Cancel saves nothing. Saving a Skill from chat works in the web app and the desktop app, where you can confirm the card. You can also go to Skills, click Add Skill, and choose Create with chat, Write skill instructions, or Upload a skill. In Agent and Project chats, Skill Creator works the same way when your role includes Create skills.
Upload either a plain Markdown file or a .zip Skill bundle with supporting files such as templates, reference documents, or scripts. You can also open an existing or imported Skill to add, upload, edit, rename, move, or delete supporting files. From chat, you can ask Langdock to copy conversation attachments into a Skill you created. Chat shows Add files to this skill? and adds nothing until you click Add files. Cancel adds nothing. If you can edit a shared Skill, asking Langdock to change it updates that Skill, including conversation files copied in the same request. If you can only view a shared Skill, chat leaves it unchanged and creates a personal copy. Each request copies up to 20 files, and a file at an existing path is replaced. The ZIP limit is 25 MB and the uncompressed limit is 20 MB. Scripts do not run on their own. When the Skill is in use, Langdock can run them in a sandbox. Disallowed file types are skipped.
Inside a ZIP archive, the main instructions and metadata file must be named SKILL.md. For a plain file upload, you can use any .md file name. You can also create and manage other Markdown files with different names in the Skill detail view.
Yes, if they’re in SKILL.md format or packaged as a .zip.
Create one skill per customer, or build one skill that contains rules for selecting the right guideline set per customer.
In a regular chat, the AI picks active Skills from their descriptions. You do not have to name the Skill in the prompt. A precise description is what makes that matching reliable. You can still choose a Skill from the @ menu when you want one on purpose.
No. Triggering is semantic and driven primarily by the skill description, not keyword matching or the skill name. Names help humans manage skills.
Skills active in normal chats apply automatically there. You can disable individual skills for a session via the skills indicator, and you can scope skills by attaching them to agents or projects.
Skills you turn on in normal chats, and skills an admin mandates there, do not carry into project chats or agent chats. Adding the Skill there is what makes it apply.
Yes. Open the Agent, find the Skills row below Instructions, and click Add skill. An attached Skill applies for people who have access to it. They do not need to install it themselves. Share the Skill with those people, a group, or the workspace. Without that access, the attached Skill does not apply.
Personal custom instructions do not apply in Agent chats. Agent instructions and attached Skills provide the instructions for those chats.
Admins can decide in Roles settings which roles can share and create skills. Only admins can share workspace-wide or force-apply skills to everyone. For individual skills, both the skill Owner and any user with Editor access can share the skill with other users and groups. Workspace admins can also share a skill with a workspace API key in Share.
Open the Skill and click Share. You can share it with people, groups, or the workspace so others can find and install it. Sharing does not turn the Skill on for them. Admins can also check Apply for all members so the Skill is required in regular chats for everyone.
Yes. Max 30 active skills per user; admins can force-apply up to 8 workspace-wide skills. Too many active skills can cause wrong-skill selection or unpredictable behavior; the limit mitigates this.
System skills can not be deactivated.
Yes. An Agent stays private until you share it. Click Share to give access to people, groups, or the workspace. If you can see that audience but cannot share with it directly, click Send request. Access stays unchanged until a reviewer approves it.
No. By default, each person uses their own connected account, such as their own Outlook inbox. Sharing the Agent does not share your login. On the Agent, you choose which actions are available and which need confirmation, within the permissions your admins allow. You can also choose a preselected connection so everyone uses the same account for that action.
Event-based triggers, such as a new email, belong in Workflows. An Agent starts when you send a prompt, or on a schedule if you create a scheduled task.
Write instructions that are specific and consistent, and include an example when you have one. Keep the Agent’s knowledge focused and test it on real cases. Models are probabilistic, so results will not be identical every time.
Use the Agents API to read an Agent’s configuration and create a copy, including in another workspace. See the Agent Migration Guide.