Skip to main content

​
Skills & memory

OrbCode can load reusable instruction sets on demand (skills) and inject project/user instructions into every system prompt (AGENTS.md memory). Both keep the model aligned with your project’s conventions without repeating yourself in every prompt.

MCP servers

Connect external tools through the Model Context Protocol

Configuration

settings.json, env vars, custom models, and sessions

​
Skills

Skills are reusable instruction sets the model can load on demand. A skill is a directory containing a SKILL.md file with optional YAML frontmatter and a markdown body of specialized instructions.

​
Creating a skill

Place a directory under ~/.orbcode/skills/ (user, applies everywhere) or .orb/skills/ (project, checked into the repo; the legacy .orbcode/skills/ location is still read for backward compatibility):
SKILL.md format:
  • description (frontmatter or first paragraph) — shown in the skill catalog.
  • when_to_use (frontmatter) — a hint telling the model when to invoke this skill.
  • ${SKILL_DIR} in the body is replaced with the skill’s absolute directory path, so you can reference bundled scripts.

​
How skills are used

The skill catalog (names + descriptions + when-to-use) is injected into the system prompt. When a task matches a skill’s when_to_use condition, the model calls the use_skill tool with the skill’s name, which loads the full instructions into context. Project skills override user skills on name collisions (closer-to-cwd wins). use_skill also accepts an explicit path in addition to a discovered name. Pass an absolute, workspace-relative, or ~/ path to a skill directory or SKILL.md file to load a shared skill that lives outside the user/project/plugin catalogs — useful from prompts or AGENTS.md @include references.

​
/create-skill — generate a skill from a description

/create-skill <describe the skill you want> turns a plain-language request into a repository-local skill under .orb/skills/<skill-name>/, keeping any supporting scripts, references, and assets inside the same skill directory. The agent is constrained to that path and cannot touch files outside it. Re-run it to refine an existing skill.

​
AGENTS.md memory

AGENTS.md files provide project- and user-level instructions that are injected into every system prompt — build commands, code style, architecture notes, conventions. This is OrbCode’s equivalent of Claude Code’s CLAUDE.md, using the open AGENTS.md filename so it works across tools. The repo-level file lives at .orb/AGENTS.md — the same folder the Orbital IDE extension reads — so a single source of truth covers your terminal and your editor. Legacy .orbcode/AGENTS.md is still read for backward compatibility.

​
Discovery

Files are loaded in this order (lowest precedence first; higher-precedence files appear later and get more weight):
  1. User memory — ~/.orbcode/AGENTS.md — personal global instructions.
  2. Project memory — AGENTS.md and .orb/AGENTS.md (and legacy .orbcode/AGENTS.md) in the cwd and every parent directory (closer-to-cwd wins). Checked into the repo, shared with the team.
  3. Local memory — AGENTS.local.md in the cwd and parents — private per-machine overrides (gitignore this).

​
@include directives

An AGENTS.md file can include other files with @path references:
@path (relative to the including file), @./path, @~/path, and @/abs/path are all supported. Includes are resolved recursively (up to 5 levels, with cycle detection).

​
/init — generate a starter .orb/AGENTS.md

The /init slash command investigates the project (directory layout, key config files, a few representative source files) and writes a concise .orb/AGENTS.md at the repo root, capped at ~150 lines so it stays cheap to include in every future prompt. The generated file covers:
  1. What the project does (1-2 lines) and its main tech stack.
  2. Project structure — the key directories/files and what each is responsible for.
  3. Architecture — how the main pieces fit together (entry points, data/control flow).
  4. Business-logic mapping — the rules and invariants the model needs to honor.
  5. Code patterns and conventions — the idioms the codebase already uses.
/init refines an existing AGENTS.md rather than overwriting it, and folds in any existing AI assistant rules (CLAUDE.md, .cursorrules, .github/copilot-instructions.md). Re-run it as the project evolves.