Skills

Skills are reusable prompt templates for workflow stages. Define a system prompt once, attach it to any LLM Stage node, and share it across workflows. Skills keep prompts versioned, consistent, and easy to update.

What Is a Skill?

A skill encapsulates:

• System prompt — The instructions sent to the LLM.
• Variables — Placeholders (e.g., {{user_query}}, {{context}}) that are filled at runtime.
• Default parameters — Optional temperature, max tokens, and stop sequences.
• Tags — Labels for organizing and searching skills.

Info: Skills are the prompt-engineering equivalent of functions. Write once, reuse everywhere.

Creating a Skill

From the Dashboard

Open Skills

Navigate to Orchestrate → Skills.

Create new

Click New Skill.

Write the system prompt

Enter the system prompt. Use {{variable}} syntax for runtime placeholders.

Set default parameters

Optionally set temperature, max tokens, and stop sequences.

Tag and save

Add tags for discoverability and click Save. The skill is now available in the Workflow Builder.

From the API

cURL

curl -X POST https://api.xilos.ai/v1/skills \
  -H "Authorization: Bearer $XILOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "refine-answer",
    "system_prompt": "You are an editor. Rewrite the following draft to improve clarity, accuracy, and tone. Preserve all factual content.\n\nDraft:\n{{draft}}",
    "variables": ["draft"],
    "default_params": { "temperature": 0.3, "max_tokens": 2048 },
    "tags": ["editing", "refinement"]
  }'

Python

import xilos
 
client = xilos.Client(api_key="...")
 
skill = client.skills.create(
    name="refine-answer",
    system_prompt=(
        "You are an editor. Rewrite the following draft to improve clarity, "
        "accuracy, and tone. Preserve all factual content.\n\nDraft:\n{{draft}}"
    ),
    variables=["draft"],
    default_params={"temperature": 0.3, "max_tokens": 2048},
    tags=["editing", "refinement"],
)
 
print(skill.id)

System Prompts

The system prompt is the core of a skill. It defines the LLM's role, behavior, and output format. Use variables to inject runtime data:

Example: Drafting Skill

You are a technical writer. Produce a clear, concise draft answering the user's question.
Use the retrieved context when available. Do not speculate beyond the provided information.

User question:
{{user_query}}

Retrieved context:
{{context}}

Example: Refinement Skill

You are an editor. Rewrite the following draft to improve clarity, accuracy, and tone.
Preserve all factual content. Do not introduce new claims.

Draft:
{{draft}}

Warning: Variable names are case-sensitive. {{User_Query}} and {{user_query}} are different placeholders.

Importing Skills

Import skills from the marketplace or from an exported JSON file.

From the Marketplace

The skills marketplace contains pre-built skills for common use cases: summarization, code review, translation, sentiment analysis, and more.

1. Navigate to Orchestrate → Skills → Marketplace.
2. Browse or search by tag.
3. Click a skill to preview its system prompt and default parameters.
4. Click Import. The skill is copied into your workspace and can be customized.

From JSON

Export a skill from one workspace and import it into another:

curl -X POST https://api.xilos.ai/v1/skills/import \
  -H "Authorization: Bearer $XILOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d @refine-answer.json

Info: Imported skills are independent copies. Updates to the marketplace original do not propagate to imported skills. Re-import to pick up changes.

Using Skills in Workflows

Attach a skill to any LLM Stage node in a workflow:

1. Open the Workflow Builder.
2. Click an LLM Stage node.
3. In the settings panel, select a skill from the Skill dropdown.
4. Map workflow variables to the skill's placeholders.
5. Optionally override the skill's default temperature and max tokens.

When the workflow executes, Xilos fills the skill's variables with runtime values and sends the assembled system prompt to the LLM.

Example Workflow

A draft-then-refine workflow uses two skills:

The Refine stage automatically receives the Draft stage's output in the {{draft}} variable.

Versioning

Skills are versioned. Every save creates a new version. Existing workflows reference the skill by ID and automatically use the latest version unless pinned. Pin a version in the workflow node settings to lock a specific version for reproducibility.

Warning: Editing a skill that is used in active workflows will change the behavior of those workflows on the next execution. Pin a version if you need stability.

Best Practices

• Keep prompts focused — One skill, one purpose. Compose multiple skills across workflow stages rather than cramming multiple behaviors into one prompt.
• Use variables — Never hardcode values that should be runtime-injected.
• Tag for discoverability — Use consistent tags so team members can find skills.
• Pin critical versions — For production workflows, pin skill versions to prevent unexpected behavior changes.
• Test before deploying — Use the Workflow Builder's Test Run to validate a skill with sample input.
• Reuse over rewrite — Check the marketplace before creating a new skill from scratch.