Skills Management

Skills are capability modules for agents. A skill is usually defined by SKILL.md, which describes when to use it, what steps to follow, which tools are available, and what output is expected.

What Are Skills

A skill is not a single tool call. It is a reusable method for a task type. For example, a "Contract Review" skill may include:

• Trigger conditions: when the user submits a contract or asks for clause review
• Steps: read the file, split clauses, identify risks, produce suggestions
• Tools: Read, Edit, WebFetch, MCP legal references
• Output format: risk summary, clause-by-clause comments, suggested edits

Dimension	Skill	Tool
Essence	High-level behavior instruction	Low-level capability interface
Granularity	A complete task workflow	One concrete operation
Example	Contract Review, Web Research, PPT Editing	Read, Grep, PowerShell, WebFetch
Contents	Instructions, steps, constraints, references, scripts	Parameters, executor, permission, result

Skill Sources

DesireCore discovers skills from several locations:

Source	Location	Scope
Global Skills	~/.desirecore/skills/	Available to all agents
Agent Skills	The agent repository's skills/ directory	Available to that agent
Project Skills	.agents/skills/ or .claude/skills/ in the work directory	Available in that project context

Disabled skills are not injected into the agent context. Skills with tool restrictions, model preferences, or provider preferences are matched against the current runtime and compute configuration.

Viewing Skills

1. Open the agent details page
2. Open Skills or Skills Management
3. Review available, disabled, and updateable skills

Skill cards can show name, description, source, version, author, status, direct-invocation setting, and tool or provider requirements. Click a skill to inspect SKILL.md, references, and scripts.

Adding Skills

From the Marketplace

1. Open the skill marketplace or the agent skills panel
2. Search or browse by category
3. Review the skill description, dependencies, and risk notes
4. Install and confirm when prompted

When installing a skill, you can choose the install target: global, a specific agent, or a project-level directory. Global marketplace skills are installed into the global skills directory where all agents can discover them. Agent-bound skills are written into the corresponding agent repository. Project-level skills only take effect in the corresponding work directory, making them ideal for maintaining team conventions, project scripts, and local templates alongside the project.

Local Import

Import Skill supports three local formats:

Format	Description
.md file	Import a single Markdown skill file; DesireCore creates the skill directory
Folder	Import a folder containing SKILL.md, scripts, references, or templates
.zip package	Import a compressed skill directory for migration or sharing

The import flow validates structure, path safety, and required fields. If it fails, the UI shows the specific reason.

Enable, Disable, Export, Remove

• Enable / Disable: controls whether a skill can be loaded. Disabling does not delete files.
• Export: in the skill detail panel, click Export to download a global skill or the current agent's embedded skill as a ZIP package. The package includes SKILL.md plus scripts, references, and assets under that skill directory, so it can be backed up, moved, or reinstalled through local import.
• Remove: deletes the skill files. Agent repository skills can be recovered from version history.
• Update: built-in and marketplace skills can detect updates. User-modified skills are not silently overwritten.

Built-in Global Skills

DesireCore includes global skills for skill creation, document processing, PDF, PPT, spreadsheet, web access, frontend design, and mail operations. Built-in skills are maintained automatically through the client and marketplace sync, while user edits are preserved where possible.

Skill Risk And Confirmation

A skill is instruction and resources; it does not bypass tool permissions. Runtime risk comes from the tools it calls, the data it touches, and the external effects it produces.

Risk Level	Meaning	Common Examples
Low	Mainly reads or organizes information without changing system state	Summarization, information lookup, format checks
Medium	Uses network access, writes files, or connects external services	Web research, report generation, API calls
High	May modify the system, run commands, send external messages, or handle sensitive data	Command execution, bulk file edits, mail sending, publishing

High-risk actions still follow tool approvals, permission settings, and human gates. Installing or enabling a skill does not approve every operation inside it; when a skill triggers sensitive tools, DesireCore applies the corresponding tool and approval policy.

Skill Metadata

SKILL.md can declare metadata in frontmatter:

Field	Description
name / metadata.name	Display name
description / metadata.description	Description used for matching
version / metadata.version	Skill version
author / metadata.author	Author
allowed-tools	Tools allowed while running this skill
provider	Preferred model or media provider slug
model	Model to use when this skill is triggered
user-invocable	Whether users can call it directly
disable-model-invocation	Whether automatic model-triggered invocation is disabled; disable_model_invocation is also accepted
context	Execution context mode, such as default or fork

These fields affect card display, command completion, automatic triggering, tool availability, and provider selection.

Custom Skill Structure

<skill_id>/
+-- SKILL.md
+-- scripts/
+-- references/
+-- assets/
+-- evals/

Clear Trigger Conditions

The clearer the skill description, the easier it is for the agent to choose it at the right time. State when to use it and when not to use it near the top of SKILL.md.

Skill Source Priority

When multiple skills are available for the same task, the agent considers them in the following order:

Source	Typical Path	Scope
Project-level Skill	<workdir>/.agents/skills/	Current project or work directory
Claude-compatible Skill	<workdir>/.claude/skills/	Compatible with existing Claude project configs
Agent Skill	agents/<agent_id>/skills/	Current agent
Global Skill	Global skills directory	All agents

When both .agents/skills and .claude/skills exist, DesireCore prefers .agents/skills. This lets you migrate to the native structure at your own pace without breaking older projects.

Skill Export

You can export individual skills from the skill detail panel. The exported ZIP includes:

• SKILL.md
• Optional scripts, references, templates, and asset files
• Metadata, input/output schemas, and localized content

Exporting skills is useful for sharing a single capability between teams rather than exporting an entire agent.

Multi-language and Resource References

Skills can declare localized content. When the UI language is set to Chinese, the agent prioritizes the Chinese description; when set to English, it loads the English version. This allows the same skill to behave consistently across multilingual teams.

Skill documents can use ${SKILL_DIR} or {{SKILL_DIR}} to reference resources within the skill directory, such as templates, scripts, or sample files. At runtime, the system resolves these placeholders to real paths, avoiding hardcoded absolute paths across different machines.

Auto-injection

Not every skill is automatically injected into the model context. Only skills that are explicitly invoked, determined to be needed by the agent, or allowed for auto-injection in their metadata are loaded. This reduces context pollution and unnecessary tool exposure.

Next Steps

• Skill Packs - Understand skills vs. tools
• Version Control - Track skill edits with Git
• Agent Marketplace - Discover more capabilities