Skills
Skills are reusable slash commands that agents and users can invoke in Claude Code. They live as markdown files in ~/.claude/commands/ and are available anywhere Claude Code runs — including inside every agent Agenties spawns.
Using skills
In the Claude Code terminal, type /skillname to invoke a skill. In Agenties chat, the orchestrator and agents can also call skills directly — they appear alongside built-in slash commands in the / autocomplete menu.
How skills are created in Agenties
Manual creation
Create a .md file in ~/.claude/commands/ with your skill definition. The file uses frontmatter for metadata and a plain text body for instructions:
markdown
---
description: Run full test suite and summarize failures
---
Run all tests with `npm test`, capture output, and write a brief summary
of failures grouped by test file. Output as markdown.
Save the file and it is immediately available as /run-tests (or whatever you named the file, without the .md extension).
Automatic creation via heartbeats
Heartbeats with the Self-improvement option enabled can analyse patterns in agent activity and generate new skill files automatically. The heartbeat agent:
→Scans recent agent sessions for repetitive multi-step tasks.
→Identifies patterns that would benefit from a reusable command.
→Writes a new .md skill file directly to ~/.claude/commands/.
→Reports the new skill name in the heartbeat summary.
Tip:Enable Self-improvement heartbeats on long-running projects. Over time, your agents will build a library of project-specific skills that make future tasks faster and more consistent.
A complete skill file looks like this:
~/.claude/commands/deploy-preview.md
---
description: Build and deploy a preview environment, then post the URL
---
1. Run `npm run build` and capture any errors.
2. If build succeeds, run `npm run deploy:preview` and capture the output URL.
3. Post the URL in a markdown code block.
4. If build fails, summarise the errors grouped by file.
Frontmatter fields
descriptionrequired
Short description shown in the autocomplete menu.allowed-tools
Comma-separated list of MCP tools this skill is allowed to use. Defaults to all tools.Best practices
Skills work best for tasks that are:
→Repetitive — you find yourself giving the same instructions multiple times.
→Multi-step — more than one or two commands to execute.
→Consistent — the same steps work every time, regardless of context.
→Delegatable — you want agents to run them without any human prompt.
Good candidates: code review checklists, deployment pipelines, test report formats, database migration runners, and environment setup scripts.
Note:Coming soon: A Skills browser in Settings to view, create, and edit skills without leaving Agenties. For now, manage skill files directly in ~/.claude/commands/.