Skip to main content

Creating Custom Skills

Build your own skills to encode your organization's standards, methodologies, and expertise into reusable agent configurations.

When to Create a Custom Skill​

Create a custom skill when:

  • Multiple agents need the same behavioral guidelines (brand voice, output format)
  • You've developed a prompt addition through iteration that you want to save and share
  • You want to onboard new team members with pre-configured best practices
  • You have a specialized domain framework that should apply broadly

Opening the Skill Editor​

  1. Click Skills in the left sidebar.
  2. Click New Skill.
  3. The Skill Editor opens.

Skill Editor Fields​

Required Fields​

Name Short, descriptive name. Clear naming helps teammates find the right skill:

  • ✅ Company Voice Writer, GDPR Compliance Checker, Internal Security Reviewer
  • ❌ My Skill, Writing, Agent Helper

System Prompt The text appended to agents using this skill. Write this as a continuation — not a replacement — of the agent's own system prompt.

Good pattern:

When producing content for [Company]:
- [Specific rule 1]
- [Specific rule 2]

Output requirements:
- [Format rule 1]
- [Format rule 2]

Avoid:

  • "You are a..." — the agent's own prompt handles persona
  • Contradicting the agent's core purpose
  • Overly long text that will be ignored (>500 words)

Optional Fields​

Description One sentence for the skill catalog: "Enforces our brand voice guidelines — conversational but professional, data-driven, customer-focused."

Category & Tags Pick a category and add tags to make the skill findable in the picker. Use existing tags where possible:

  • content, marketing, social, research, data, developer, devops, cybersecurity, customer-success, sales, finance, legal, hr, design, operations, product, education

Example Prompts 2-4 sample user messages that demonstrate the skill working well. These appear in the skill preview when teammates browse the catalog.

"Write a blog post about our new API rate limiting feature"
"Create a social media post announcing our Series B"
"Draft the release notes for v2.4.0"

Version A version string for your own tracking: 1.0.0. Increment when you make significant changes.

Prompt-only

Skills carry only a system_prompt (plus metadata for browsing). There are no model or temperature settings on a skill — the agent always controls its own model and temperature.

Example: Company Voice Writer Skill​

Name: company-voice-writer
Tags: writing, marketing, brand

System Prompt:
Writing standards for Acme Corp content:

Voice and Tone:
- Conversational but professional — write how a smart colleague talks, not how a lawyer writes
- Lead with customer outcomes, not features ("You'll ship faster" not "Our CI/CD system is fast")
- Use "you" and "we" — avoid passive voice and impersonal constructions
- Confidence without arrogance — state what we know, acknowledge what we don't

Vocabulary:
- Product names: Acme Platform (not "the platform"), Acme API (not "our API")
- Avoid: "leverage", "synergy", "robust", "seamless", "cutting-edge", "game-changing"
- Preferred: "use", "build", "fast", "reliable", "powerful"

Numbers and Data:
- Lead with the specific number: "3x faster" not "significantly faster"
- Source claims: include "[Source: Internal benchmark, Q3 2024]" for statistics
- Round to meaningful precision: "200ms" not "194.7ms"

Example Prompts:
- "Write a product launch email for our new webhooks feature"
- "Create a landing page headline for the enterprise tier"
- "Write 3 tweet options announcing our GitHub integration"

Example: Security Review Checklist Skill​

Name: internal-security-reviewer
Tags: security, code, engineering

System Prompt:
Security review process for Acme Corp codebase:

Mandatory checks (every review):
1. Authentication: All endpoints authenticated? Session tokens properly validated?
2. Authorization: RBAC enforced? Privilege escalation paths closed?
3. Input validation: All user input sanitized? SQL parameterized? Output encoded?
4. Secrets: No API keys/passwords in code? Env vars properly handled?
5. Dependencies: Known CVEs in new packages? (reference NIST NVD)

Output format:
## Security Review: [Component Name]

### 🔴 Blockers
[Must fix before merge]

### 🟡 Warnings  
[Should fix soon]

### ✅ Passing
[What looks good]

### 📋 Compliance Notes
[GDPR/SOC2 relevant observations]

Sign off: "Reviewed by AI security checker — human review required for Blockers"

Saving and Sharing​

Click Save Skill. The skill is immediately available to all project members:

  • Appears in the Skills list with your username as creator
  • Available in the skill picker in the Agent Editor
  • Shown with custom tag in the catalog (distinguishable from built-in skills)

Updating Skills​

  1. Click Skills in the sidebar.
  2. Click the skill to open the editor.
  3. Make changes.
  4. Click Save Skill.

All agents using this skill receive the updated prompt on their next run. Changes take effect immediately — there's no deploy step.

Test after updating skills

When you update a skill, all agents using it change behavior. Test the updated skill in the agent's test console before running production workflows.

Deleting a Skill​

  1. In the Skills list, click the three-dot menu on the skill.
  2. Click Delete.
  3. Confirm.

When a skill is deleted, agents that had it assigned lose the skill. Their system prompts revert to just the agent's own prompt. No data is lost — the skill is simply no longer applied.

Skill Governance Tips​

Version Your Skills​

Include a version number in the skill name or description: company-voice-writer v2. This makes it clear which version agents are using and when you've made significant changes.

Document What Changed​

When updating a skill, add a comment to the description: Updated Jan 2024 — removed "leverage" from avoid list, added GDPR compliance reminder.

Test Before Deploying​

Before updating a production skill, create a copy (company-voice-writer-test), update the copy, assign it to a test agent, run through your test cases, then update the production skill.

Limit Skill Scope​

Skills work best when they're focused. A skill that tries to do everything becomes a second system prompt and creates conflicts. Create separate skills for separate concerns:

  • company-voice — just writing style
  • company-output-format — just output format requirements
  • company-legal-disclaimers — just legal requirements

Assign multiple skills if needed... actually, each agent only gets one skill. Combine focused concerns into a single coherent skill.

Last updated on