Skip to main content

Skills and memory

A skill is a folder with a SKILL.md file: a reusable instruction set the AI can pull in on demand (agentskills.io spec). Memory stores durable per-user facts across chats. And if you already keep skills or a CLAUDE.md/AGENTS.md for another tool, Computer picks them up automatically.

Skill format​

---
name: release-checklist
description: How we cut a release in this repo
---

1. Run the full test suite...

Frontmatter needs name and description; the body is the instructions. Optional subdirectories: references/, templates/, scripts/, assets/. Skills load progressively: the catalog is always visible to the model, the full SKILL.md loads on activation, and resources load on demand.

Where skills are discovered​

ScopePaths
Workspace.cptr/skills, .agents/skills, .claude/skills, .codex/skills
Global~/.cptr/skills, ~/.agents/skills

Skills you already have for Claude Code or Codex are picked up automatically, with no copying.

Using and managing skills​

  • Type $ in the chat input to mention a skill; / also finds skills.
  • Skills created in the UI are written to .cptr/skills (workspace) or ~/.cptr/skills (global).
  • The admin Skills page controls which skills are enabled.
  • Optional background skill review: every N turns (default 10), the AI can propose updates to your skills based on how the work actually went.

Memory​

Memory is per-user: toggle it in Settings. Recalled facts are included in later chat context. Entries live under <data-dir>/memory/users/<id>/ and can be reviewed, edited, or deleted from the Memory view. Don't put secrets in skills or memory; both end up in model context.

System prompts​

The system prompt resolves in order; the first match wins:

  1. Workspace file <workspace>/.cptr/system.md
  2. Per-model system prompt (model settings)
  3. Global (the * model)
  4. Built-in default

Prompts support {{VAR}} template variables: WORKSPACE_NAME, WORKSPACE_PATH, FILE_TREE, INSTRUCTIONS, MEMORY, SKILLS, CPTR_CONTEXT, RUNTIME_ENV, HOSTNAME, OS, PLATFORM, ARCH, SHELL, HOME, CPTR_VERSION, DATE, MODEL.

Auto-loaded instruction files​

Computer reads instruction files from the workspace root in this order (up to 32KB): MEMORY.md, AGENTS.md, AGENT.md, CLAUDE.md. An existing CLAUDE.md or AGENTS.md just works: the AI sees it on every turn, alongside a compact file tree (which skips .git, node_modules, .venv, dist, ...), memory, and the skills catalog.

This content is for informational purposes only and does not constitute a warranty, guarantee, or contractual commitment. Open WebUI is provided "as is." See your license for applicable terms.