Skills

Skills carry reusable experience assets in Flocks, including standards, policies, manuals, methodologies, checklists, and task templates. If tools define executable actions, Workflow defines stable processes, and Agent defines roles and responsibilities, then a Skill defines the method an Agent should follow in a specific scenario.

A Skill is not a script, tool, or Agent. It is a methodology asset loaded into the Agent runtime. It usually centers on SKILL.md and can include scripts, reference files, or templates. When an Agent handles a certain type of task, it can load the corresponding Skill to make reasoning, inspection order, and output format more stable.

1. Purpose

1.1 What Skill Solves

A Skill defines the method a class of tasks should follow for judgment and execution. It is suitable for preserving experience, standards, and methods, rather than fixing every action step.

A typical Skill can include:

Applicable scenarios: when this Skill should be loaded.
Methodology: inspection order, judgment criteria, reasoning framework, and notes.
Output format: report structure, risk levels, evidence list, and handling recommendations.
Dependency declarations: the body can describe required tools, MCP, scripts, or external environments. Current structured availability checks only recognize dependencies such as commands and environment variables.
Auxiliary files: references, templates, sample data, or read-only scripts.

1.2 Relationship with Other Modules

Module	Relationship
Session Management	Users can ask Rex in a session to load, create, audit, or adjust Skills.
Agent	Agents are the main loaders of Skills; Skills affect Agent reasoning and action selection.
Workflow	Workflow fixes process, while Skill provides methodology. Workflow itself has no separate Skill configuration field; the Agent executing the task can load Skills according to its own configuration or delegation parameters.
Tools	Skill body text can describe tool or MCP requirements; current automatic availability checks only recognize commands, environment variables, and installation metadata.
Workspace	Skills can guide Agents on how to use materials and outputs in Workspace. Skill files themselves are stored in the user directory or current project directory and are not managed by the Workspace module.

1.3 Built-In Platform Skills

Flocks also uses Skills to carry platform-building capabilities. Typical built-in Skills in the current codebase include tool-builder, workflow-builder, agent-builder, skill-builder, browser-use, and webui-page-builder. Some system Skills have category: system or ui_hidden: true and do not appear in the normal Skills management list.

2. Use Cases

2.1 Good Fits for Skill

Preserve judgment methods: alert triage, malware sample analysis, vulnerability impact assessment, host investigation checklists.
Unify output format: make different Agents output with the same report structure.
Reuse team experience: convert expert experience, standards, and handling manuals into loadable assets.
Flexible task paths: not suitable to hard-code as a Workflow, but still requiring stable inspection order and judgment criteria.
Multiple Agents need the same method: for example, several security Agents all load the same IOC triage method.

2.2 Poor Fits for Skill

If you need to execute a fixed sequence of actions, prefer Workflow.
If you need a professional execution role for a class of tasks, prefer Agent.
If it is a one-off temporary question, hand it directly to Rex in a session.

Simple rules:

If the next time is about how to judge, create a Skill.
If the next time should automatically run the same steps, create a Workflow.
If the next time should be assigned to a fixed role, create an Agent.

3. Ways to Create and Install Skills

3.1 Option 1: Install from the Page Button

On the Skills page, click Install Skill and enter the Skill source. The page currently supports source forms including:

text

clawhub:github
github:owner/repo
safeskill://...
skills-sh:owner/repo/skill-name
https://raw.githubusercontent.com/...
/local/path/to/skill

The WebUI currently does not provide scope selection. Page installation and page creation both write to the user-level directory ~/.flocks/plugins/skills/ by default. A local path refers to a path on the machine running the Flocks service, not the computer where the remote browser is running.

After installation, the page refreshes the Skill list. If the Skill declares dependencies that the current environment does not satisfy, the page shows the missing dependencies and can continue with the "install dependencies" operation.

3.2 Option 2: Install from the Command Line

You can also install Skills through the Flocks CLI:

bash

flocks skills install clawhub:github
flocks skills install skills-sh:owner/repo/skill-name
flocks skills install safeskill://...
flocks skills install github:owner/repo
flocks skills install https://raw.githubusercontent.com/...
flocks skills install /local/path/to/skill

If a repository contains multiple Skills, use --skill to specify the subdirectory:

bash

flocks skills install github:owner/repo --skill skill-name

The default target is the user-level directory. To install into the current project directory, specify scope:

bash

flocks skills install github:owner/repo --scope project

The project-level installation location is .flocks/plugins/skills/ under the current project when the command is executed. It is unrelated to the Workspace business file directory.

After installation, check status and install dependencies:

bash

flocks skills status
flocks skills install-deps <skill-name>

Common commands:

Command	Purpose
`flocks skills list`	View installed Skills.
`flocks skills status`	View Skill status and dependency information.
`flocks skills find <query>`	Search installable or known Skills.
`flocks skills install <source>`	Install a Skill from clawhub, GitHub, URL, or local path.
`flocks skills remove <name>`	Remove a Skill from the user-level directory; project-level and built-in Skills cannot be removed.
`flocks skills install-deps <skill-name>`	Install dependencies declared by a Skill.

If using the npm wrapper, also use the plural command group. The npm wrapper passes through to the Python flocks CLI and does not rewrite singular skill into plural skills:

bash

npx @flocks-ai/flocks skills install clawhub:github
npx @flocks-ai/flocks skills install github:owner/repo

In the Flocks Python CLI, the official command group is plural: flocks skills ....

3.3 Option 3: Ask Rex to Install from an External Source

You can install Skills from clawhub, a URL, GitHub, or an internal team repository. For example, say to Rex in a session:

text

Install <skill-name> from clawhub

Or:

text

Install this Skill from <URL> and check whether dependencies are satisfied.

Before installing external Skills, perform a security review. Pay special attention to Prompt injection, hidden scripts, outbound network traffic, hard-coded credentials, and misleading capability descriptions.

3.4 Option 4: Install from a Local Directory

Both CLI and WebUI can use a local directory containing SKILL.md or a SKILL.md file as the installation source:

bash

flocks skills install /path/to/my-skill

The current local-path installer only reads and saves SKILL.md. It does not copy auxiliary files such as scripts/, references/, or templates/ in the same directory. If the Skill depends on those files, manually copy the complete directory to the target location, or use a GitHub/package source that can download the complete directory.

3.5 Option 5: Ask Rex to Create a Skill

You can ask Rex directly in a session:

text

Help me create a Skill for "detecting malicious Skills."

Goal: scan Skills downloaded from clawhub and other external sources to determine whether security risks exist.

Inspection areas:
- Abnormal characters / Prompt injection
- Executable commands / Bash command execution
- Network connections / data exfiltration
- Capability description inconsistent with actual behavior
- Base64 or other obfuscation
- Hard-coded IP / password / C2 address
- Guidance that tells Agents to ignore safety guardrails

Output: risk level (Critical / High / Medium / Safe) + evidence list + handling recommendations

3.6 Option 6: Promote a Successful Task into a Skill

When Rex or an Agent has successfully completed a task and its judgment method is worth reusing, continue with:

text

Based on the triage process just now, help me preserve a reusable Skill.
Focus on keeping the judgment criteria, inspection order, evidence requirements, risk grading, and output format.

This is suitable for converting personal experience or a successful handling case into a reusable methodology asset for the team.

4. WebUI Flow

4.1 Open the Skills Page

Go to Agent Studio -> Skills from the sidebar. The page usually shows:

Project-level, user-level, built-in, and compatibility-source Skills discovered in the current request context.
Discovery source labels for Skills, such as project, user, flocks, and claude.
Enablement status and dependency satisfaction.
SKILL.md content and dependency information.
Search, source filter, enable/disable, refresh, install, create, detail view, and dependency installation entry points.

User-level Skills can be modified and deleted through API/WebUI. Project-level Skills are read-only in WebUI and should be maintained directly in the project directory through version control.

4.2 Rex Generates a Skill

When creating a Skill through natural language, Rex usually:

References the directory structure and SKILL.md format of existing Skills.
Creates a new Skill directory.
Generates SKILL.md.
Generates helper scripts, templates, or reference files as needed.
Checks dependency tools and usage boundaries.
Presents the new Skill in the Skills list.

SKILL.md usually needs to be fairly complete, so generation can take longer than creating a simple Agent.

4.3 Load and Use a Skill

In a session, you can explicitly ask Rex or an Agent to load a Skill:

text

Use Skill Security Audit to audit the NakedIn and SystemOptimizer Skills.

After the Skill is loaded, the Agent follows the methodology, inspection order, and output format in SKILL.md.

4.4 Dynamically Execute a Skill through Category Delegation

Rex can use delegate_task(category=..., load_skills=[...]) to hand a task to a category executor. Category delegation currently uses Rex-junior and appends the full content of the selected Skill to that subtask's prompt. It does not replace Rex-junior's own System Prompt and cannot be specified directly through subagent_type="rex-junior".

This mode is suitable for letting an independent subtask follow a specific methodology. If a Skill is frequently executed through category delegation and its boundaries and tool set gradually stabilize, promote it into a formal Agent.

4.5 `/skills` Command in Sessions

In WebUI sessions, TUI, or CLI sessions, use /skills directly to view currently available Skills.

Command	Purpose
`/skills` or `/skills list`	List currently available Skill names and descriptions.
`/skills refresh`	Rescan and refresh the Skill list.

For example:

text

/skills
/skills refresh

If an Agent does not load a Skill as expected, first use /skills to confirm that the Skill has been recognized by the system, then check whether the SKILL.md name, description, and applicable scenarios are clear.

4.6 View and Install Skills in TUI

TUI supports viewing the Skill list in sessions and provides a Skill installation dialog. Common usage includes:

Use /skills in the input box to view available Skills.
Open the Skill dialog to view Skill descriptions, paths, and dependency status.
Install Skills from clawhub:, github:, URL, or local path through the installation dialog.
Select global or project scope in the installation dialog.
If dependencies are missing after installation, run install-deps.

The TUI Skill dialog is suitable for everyday browsing and selection. Batch installation, dependency checks, and automation scenarios are better handled with CLI.

5. Skill File Structure and Installation

5.1 Storage Locations

Flocks discovers Skills from multiple compatible locations. Common installable locations are:

Type	Storage Location	Description
Project-level Skill	`.flocks/plugins/skills/<name>/SKILL.md` under the current project	Follows project files and version control; discovered in that project's request context and treated as a read-only resource by WebUI/API.
User-level Skill	`~/.flocks/plugins/skills/<name>/SKILL.md`	Reusable by the current operating system user; objects created, modified, and deleted by WebUI live here.
Built-in Flocks Skill	`.flocks/plugins/skills/` or compatible `.flocks/skill[s]/` in the Flocks installation directory	Provided with product code and should not be modified directly from the Skills page.
Claude-compatible Skill	Global or project `.claude/skills/`	A compatibility discovery source with the lowest priority.

The project directory means the current request directory corresponding to Instance.get_directory(). Discovery also scans upward from that directory to the current worktree root, so this is not the Workspace business file directory shown in WebUI.

Multiple sources can coexist. The code handles duplicate Skill names by letting later scans override earlier scans. The simplified priority is:

text

User-level > current project-level > Flocks built-in/provided with source > Claude-compatible directories

Therefore user-level Skills override project-level Skills of the same name, which is the opposite of the order in older documentation. In practice, avoid duplicate names.

5.2 Directory Structure

Each Skill is an independent folder:

text

skills/
└── skill-security-audit/
    ├── SKILL.md
    ├── references/
    ├── templates/
    └── scripts/

Where:

SKILL.md is the main file and defines the Skill's goal, applicable scenarios, methodology, dependencies, and output format.
references/ can store reference materials, standards, examples, and explanatory documents.
templates/ can store output templates, report templates, or Prompt templates.
scripts/ can store helper scripts, but their purpose and security boundaries should be explicit.

5.3 Install a New Skill

Place a new Skill folder under .flocks/plugins/skills/ or ~/.flocks/plugins/skills/ to install it. The system recognizes its SKILL.md and loads it into the Skills list or runtime.

When installing an external Skill, do not only check whether the description is useful. Review the scripts, templates, and reference files in the full directory.

5.4 SKILL.md Metadata and Dependencies

SKILL.md must start with YAML frontmatter and provide valid name and description; otherwise the discoverer ignores the Skill. name must match [a-z0-9]+(-[a-z0-9]+)*, and description must be 1 to 1024 characters long.

Minimal example:

yaml

---
name: ndr-alert-triage
description: Triage NDR alerts using a fixed evidence chain and output conclusions.
---

Skill names use lowercase kebab-case, for example:

text

skill-security-audit
ndr-alert-triage

If a Skill depends on external tools or runtime environments, declare them explicitly. Flocks' dependency installation flow can handle several dependency types:

Dependency Type	Suitable For
`brew`	macOS system packages or command-line tools.
`npm`	Node.js ecosystem commands or packages.
`uv`	Python project environments and package management.
`pip`	Python packages.
`go`	Go language tools.
`download`	Download binaries or archives from declared URLs.

Dependency metadata is read from metadata.flocks and is compatible with metadata.openclaw. Runtime conditions support requires.bins, requires.any_bins, and requires.env; installation declarations use the install list. The clearer the dependency declarations, the easier it is for the page and CLI to show "what is missing" and "how to install it." For production Skills, run a dependency check and a real task validation after installation.

6. Run, Verify, and Adjust

6.1 Verify Whether a Skill Meets Expectations

Use 1 to 2 real tasks to validate the Skill and focus on:

Whether the Agent can correctly identify when to load the Skill.
Whether the applicable scenarios in SKILL.md are clear.
Whether the inspection order and judgment criteria are stable.
Whether the output structure meets team requirements.
Whether dependency tools or scripts are available and safe.

6.2 View Outputs

Skills do not directly produce results. Outputs are generated by the Agent, Rex, or category-delegated subtask that loads them. Common outputs include:

Structured conclusions in the session.
Markdown reports.
Evidence lists and risk grading.
Files under Workspace outputs/.

6.3 Adjust a Skill

If a Skill does not perform well, clearly state in a Rex session which Skill to optimize and which capability needs improvement.

For example:

text

Help me optimize skill-security-audit.
The current Prompt injection judgment is too broad. Add evidence references, risk level boundaries, and false-positive exclusion conditions.

Rex can adjust SKILL.md, output templates, reference files, or helper scripts based on feedback. Validate again with real samples after adjustment.

7. Core Concepts

7.1 What Is SKILL.md?

SKILL.md is a methodology document for the Agent runtime. It usually includes:

Goals and applicable scenarios.
Non-applicable scenarios.
Judgment criteria.
Checklists and step order.
Output format conventions.
Required tools or MCP.
Security boundaries and notes.

After an Agent loads a Skill, it uses this document as methodological guidance, affecting reasoning and action selection.

7.2 Skill vs Agent vs Workflow

Dimension	Skill	Agent	Workflow
What it is	Methodology / standard / task template	Role + tools + Prompt + execution mode	Executable process / node graph
Can it execute independently?	No, it must be loaded by an Agent	Yes, it can have its own loop	Yes, the platform runs it according to the graph
Best at	Soft judgment, checklists, experience preservation	Deep handling of a class of tasks	Stable steps, batch processing, scheduled runs
How it is generated	Created through natural language or promoted from experience	Created through natural language or promoted from tasks	Created through natural language or promoted from processes

7.3 Skill Loading and Priority

Multiple directories can contain Skills at the same time, including Claude-compatible directories, built-in directories, project-level directories, and user-level directories.
For duplicate names, user-level takes priority over project-level, and project-level takes priority over lower-priority built-in and Claude-compatible sources.
Disabled Skills still appear in the management list, but they do not appear in the available Skill summary for Agents and cannot be loaded through skill_load or delegation parameters.
When commands or environment variables declared in metadata.flocks.requires are not satisfied, the Skill is marked unavailable and missing dependencies are shown. Tool and MCP requirements written only in the body must be confirmed manually before use; the system does not automatically determine availability from them.
The clearer the task description, the easier it is for the Agent to load the correct Skill.

7.4 Installation Risks

Skills downloaded from external sources may introduce security risks. Typical risk patterns include:

Prompt injection: asking the Agent to ignore system instructions or override security rules.
Hard-coded C2 exfiltration: suspicious outbound addresses or data exfiltration logic in scripts.
Misleading capability description: described as a normal tool but actually executing dangerous commands.
Hidden script execution: reading system files, modifying configuration, or downloading binaries.
Base64 obfuscation: hiding real behavior through obfuscation.
Social engineering guidance: language that encourages the Agent to bypass restrictions.

Review external Skills manually, or run a dedicated audit Skill first, before deciding whether to install them in production.

7.5 Skill Capability Boundaries

Skills are suitable for preserving methods, not for carrying all execution logic. In practice:

Keep one Skill focused on one methodology.
Write applicable and non-applicable scenarios clearly.
Make output formats concrete; avoid only abstract principles.
Keep scripts minimal and preferably read-only.
Declare external dependencies explicitly and avoid implicit calls.

8. Real Example: Malicious Skill Audit

8.1 Background

The user wants to create a Skill for detecting malicious Skills, scanning skill packages downloaded from clawhub and other external sources to determine whether security risks exist.

8.2 Creation Process

In a Rex session, the user describes detection directions such as abnormal characters, Prompt injection, command execution, network connections, data exfiltration, misleading capability descriptions, Base64 obfuscation, hard-coded IPs / passwords / C2 addresses, and so on.

Rex:

References the format of existing Skills.
Creates the skill-security-audit directory.
Generates SKILL.md.
Designs inspection stages such as file collection, static detection, semantic detection, and overall judgment.
Generates helper review scripts as needed.

8.3 Audit Flow

Ask Rex in a session to use the Skill to audit target Skills:

text

Use skill-security-audit to audit the NakedIn and SystemOptimizer Skills.

The execution usually includes:

Loading skill-security-audit.
Reading the full directories of the target Skills.
Checking SKILL.md, scripts, templates, and reference files.
Outputting risk levels, evidence lists, and handling recommendations.

8.4 Result Evaluation

This example shows that Skill creation can start from a high-level requirement, then Rex fills in inspection logic, output structure, and execution steps. Before production use, still perform a manual review to confirm that the methodology is accurate, risk grading is reasonable, and the output format meets team requirements.

9. FAQ

9.1 I Installed a Skill, but the Agent Does Not Call It as Expected. What Should I Do?

First make the task goal and desired Skill more explicit. If it still does not trigger, check whether the SKILL.md name, description, applicable scenarios, and trigger conditions are clear.

9.2 Does Installing a Skill Only Load SKILL.md?

If the Skill depends on scripts, templates, or references, keep the entire Skill folder. GitHub or archive installation can include these auxiliary files. The current local-path installation only writes SKILL.md; copy the rest manually.

9.3 How Should I Organize a Successful Operation into a Skill?

Clarify where to start next time a similar task appears, which signals determine risk level, what evidence should be collected, and what structure the output should have.

9.4 Why Do Downloaded Skills Have Security Risks?

Skills from community or external sources may contain Prompt injection, hidden scripts, data exfiltration, or misleading capability descriptions. Audit the full directory before installation, not just the name and summary.

9.5 Can a Skill Include Executable Scripts?

Yes, but scripts are also one of the main risk sources. Publicly shared Skills should avoid bundled scripts when possible. Internal Skills should limit script scope and clearly define read-only or allowlist constraints.

9.6 How Do I Distinguish Skill from Agent Prompt?

Agent Prompt: defines the long-term responsibilities, capability boundaries, and execution style of a role.
Skill: defines the methodology for a class of tasks. It can be loaded by multiple Agents or injected into an independent subtask through category delegation parameters.

Session Management: Create, load, audit, and adjust Skills through Rex.
Agent: The main loader of Skills.
Workflow: Workflow and Skill respectively carry "process" and "method."
Tools: Skill body text can record tool requirements, while structured dependency checks focus on commands and environment variables.
Workspace: Skill files can be stored at user level or current project level, but are not managed by Workspace.
Scenario Practice - Threat Intelligence and IOC Triage: Turn IOC triage methods into Skills.
Scenario Practice - Alert Triage: Preserve alert triage methodology as a Skill.

Operation demo videos will be added to this page after release materials are stable.

Skills

1. Purpose

1.1 What Skill Solves

1.2 Relationship with Other Modules

1.3 Built-In Platform Skills

2. Use Cases

2.1 Good Fits for Skill

2.2 Poor Fits for Skill

3. Ways to Create and Install Skills

3.1 Option 1: Install from the Page Button

3.2 Option 2: Install from the Command Line

3.3 Option 3: Ask Rex to Install from an External Source

3.4 Option 4: Install from a Local Directory

3.5 Option 5: Ask Rex to Create a Skill

3.6 Option 6: Promote a Successful Task into a Skill

4. WebUI Flow

4.1 Open the Skills Page

4.2 Rex Generates a Skill

4.3 Load and Use a Skill

4.4 Dynamically Execute a Skill through Category Delegation

4.5 /skills Command in Sessions

4.6 View and Install Skills in TUI

5. Skill File Structure and Installation

5.1 Storage Locations

5.2 Directory Structure

5.3 Install a New Skill

5.4 SKILL.md Metadata and Dependencies

6. Run, Verify, and Adjust

6.1 Verify Whether a Skill Meets Expectations

6.2 View Outputs

6.3 Adjust a Skill

7. Core Concepts

7.1 What Is SKILL.md?

7.2 Skill vs Agent vs Workflow

7.3 Skill Loading and Priority

7.4 Installation Risks

7.5 Skill Capability Boundaries

8. Real Example: Malicious Skill Audit

8.1 Background

8.2 Creation Process

8.3 Audit Flow

8.4 Result Evaluation

9. FAQ

9.1 I Installed a Skill, but the Agent Does Not Call It as Expected. What Should I Do?

9.2 Does Installing a Skill Only Load SKILL.md?

9.3 How Should I Organize a Successful Operation into a Skill?

9.4 Why Do Downloaded Skills Have Security Risks?

9.5 Can a Skill Include Executable Scripts?

9.6 How Do I Distinguish Skill from Agent Prompt?

10. Related Modules

4.5 `/skills` Command in Sessions