Skills

A Skill is a reusable set of instructions you can attach to one or more Assistants (internally “Agents”). Each Skill is a SKILL.md file: Markdown that teaches an Assistant how to approach a task — a code reviewer’s checklist, a writing coach’s editing rules, a data analyst’s rigor. A Skill can also bundle a Connection (an MCP server config) so the Assistant gains extra tools when the Skill is enabled.

Skills keep behavior modular. Instead of cramming every instruction into one Assistant’s Personality (internally “Persona”), you compose capabilities by stacking small, focused Skills — and reuse the same Skill across many Assistants.

How Skills are applied

When an Assistant starts (internally “spawns”), RondoFlow loads every Skill that is enabled for that Assistant, ordered by priority (ascending). Each enabled Skill’s SKILL.md body is collected under a ## Skill: <name> heading and forwarded to the spawned Claude Code CLI through its own flag — --append-system-prompt — which is separate from the Persona, which is passed via --system-prompt. Keeping the two apart means a Skill’s instructions never overwrite the Personality; they’re layered on top of it. If a Skill carries a Connection (MCP) config, that server is also registered for the run, so its tools become available alongside any other Connections.

The append-prompt mechanism is also used for Assistants running on API providers (OpenAI, Perplexity). Those providers run over an HTTP API rather than the Claude Code CLI, so the Persona and the enabled Skills are combined into the request’s system prompt — Skills still apply — while the Claude-CLI-only concerns (model mapping, allowed-tool lists, MCP registration, external folders) are skipped. See Connections for the provider differences in detail.

Built-in catalog

RondoFlow ships a large built-in catalog of ready-to-install Skills — 2,300+ entries spanning categories like AI & Agents, Finance & Crypto, Development, Data & Analytics, Media, Productivity, DevOps & Infra, Content & Writing, and more. A handful are first-party (authored by rondoflow); the vast majority are contributed by the community. The catalog ships in the @rondoflow/catalog package.

Browse it in the Skill Marketplace panel (open it from an Assistant to install and attach in one step, or stand-alone to just install). The Marketplace derives its category chips from the loaded catalog — each chip shows the category and its count — and a search box matches a Skill’s name, description, or category.

The five first-party Skills make good starting points:

A Skill’s author is rondoflow for the first-party entries (all at version 1.0.0) and community for contributed ones. The catalog’s icon values are lucide icon names used in the UI — they are not stored on installed Skills.

Installing Skills

There are three ways to add a Skill. All of them write to ~/.rondoflow/skills/<name>/ and register a record in the database.

Requirements
• A SKILL.md file at the repository root (or a lowercase skill.md)
• The repository must be publicly accessible (or credentials configured)
• An optional rondoflow.json may override SKILL.md frontmatter and declare an MCP (Connection) config

Skill metadata

Every Skill stores the following fields, parsed from the SKILL.md frontmatter (and merged with rondoflow.json, which wins on conflicts):

mcpConfig describes a stdio MCP server bundled with the Skill. The common shape is a single server object:

{
  "type": "stdio",
  "command": "my-mcp-server",
  "args": ["--flag", "value"],
  "env": { "API_TOKEN": "..." }
}

You can also declare multiple servers by storing mcpConfig as a map of server name → config (any object without a top-level command key is treated as a map):

{
  "search": { "type": "stdio", "command": "search-mcp", "args": [] },
  "fetch":  { "type": "stdio", "command": "fetch-mcp",  "args": [] }
}

When a Skill carries an MCP config, its servers are registered for any Assistant that has the Skill enabled, namespaced to avoid collisions: a single-object config registers as <skillName>/default, and each entry in a map registers as <skillName>/<serverName>. See Connections for how MCP servers are wired into a run.

On-disk layout

Installed Skills live under your home directory:

• ~/.rondoflow
  • skills
    • code-review
      • SKILL.md
    • my-custom-skill
    • imported-skill

A valid SKILL.md is YAML frontmatter (at least name and description) followed by the instruction body:

---
name: code-review
description: Reviews code for quality, bugs, security issues, and best practices
category: Development
author: rondoflow
version: 1.0.0
---

# Code Review Skill

You are an expert code reviewer...

Attaching a Skill to an Assistant

A Skill does nothing until it is attached to an Assistant. Each attachment has two controls:

• Priority — an integer (default 0) that sets the order Skills are applied. Lower numbers apply first. When you install or add a Skill from the Marketplace in attach mode, it gets max + 1 so it lands at the bottom of the existing order.
• Enabled — a toggle (default on). Disabled Skills stay attached but are skipped at run time, so you can mute a Skill without removing it.

There are two ways to attach a Skill:

Conflict detection

When an Assistant has two or more enabled Skills, RondoFlow runs heuristic conflict detection and surfaces warnings. It checks each pair of enabled Skills (ordered by priority) for:

These are advisory, not blocking — they help you spot incompatible or redundant Skill combinations (for example, two same-category Skills whose advice pulls in different directions, or two Skills bundling the same Connection). Resolve them by disabling one Skill, adjusting priority, or removing a duplicate. Separately, the prompt builder logs a warning at spawn time when two MCP servers with the same command + args signature come from different sources.

Skills API

Skills are managed over the REST API. Responses use the standard { success, data?, error?, meta? } envelope.

Related