The Problem: Manual Skill Invocation

Every session, you type /content-seed to activate the harvesting skill. Every session, you forget at least once. The skill exists, but it’s not part of your workflow - it’s an interruption to it.

What if skills could be always-on? Loaded automatically when the agent starts, running passively in the background?

The Solution: Skill Injection via Config

flowchart LR
    subgraph CONFIG["oh-my-opencode.json"]
        AGENT["Agent Config"]
        SKILLS["skills: ['content-seed']"]
        AGENT --> SKILLS
    end
    
    subgraph STARTUP["Agent Startup"]
        LOAD["Load Agent"]
        INJECT["Inject Skills"]
        READY["Ready with Skills"]
        LOAD --> INJECT --> READY
    end
    
    CONFIG --> STARTUP
    
    style SKILLS fill:#22c55e,color:#000
    style READY fill:#3b82f6,color:#fff

In OpenCode (and similar agent frameworks), you can specify skills that auto-load for specific agents:

// ~/.config/opencode/oh-my-opencode.json
{
  "agents": {
    "sisyphus": {
      "model": "anthropic/claude-sonnet-4-20250514",
      "skills": ["content-seed", "git-workflow"]
    },
    "oracle": {
      "model": "anthropic/claude-opus-4-20250514",
      "skills": ["deep-analysis"]
    }
  }
}

Now when sisyphus agent starts, it automatically has the content-seed and git-workflow skills active. No manual /skill commands needed.

Why This Matters

Before: Manual Invocation

User: starts session
User: /content-seed
Agent: Content seed skill loaded
User: actually does work
... forgets to harvest seeds at end ...

After: Always-On

User: starts session
Agent: (content-seed already active, harvesting passively)
User: does work
Agent: (automatically captures content seeds)

The skill becomes a behavior, not a command.

Use Cases

1. Passive Harvesters

Skills that watch for interesting patterns and capture them:

{
  "skills": ["content-seed", "error-collector", "pattern-detector"]
}

These run in the background, noting things worth saving without interrupting your flow.

2. Workflow Enforcers

Skills that ensure certain practices are followed:

{
  "skills": ["git-workflow", "test-reminder", "doc-checker"]
}

The agent always checks for tests, always follows git conventions, always reminds about documentation.

3. Context Providers

Skills that provide specialized knowledge:

{
  "skills": ["project-context", "api-reference", "style-guide"]
}

The agent always knows your project’s conventions, API patterns, and style preferences.

Configuration Patterns

Per-Agent Skills

Different agents need different skills:

{
  "agents": {
    "sisyphus": {
      "skills": ["content-seed", "git-workflow"]
    },
    "oracle": {
      "skills": ["deep-analysis", "architecture-patterns"]
    },
    "explore": {
      "skills": ["codebase-mapping"]
    }
  }
}

Skill with Parameters

Some frameworks support skill configuration:

{
  "agents": {
    "sisyphus": {
      "skills": [
        "content-seed",
        {
          "name": "git-workflow",
          "config": {
            "requireIssue": true,
            "maxCommitsPerBranch": 7
          }
        }
      ]
    }
  }
}

Skill Sources

Skills can come from multiple locations:

{
  "skills": {
    "sources": [
      { "path": "~/.config/opencode/skills", "recursive": true },
      { "path": "~/Dev/my-skills/skills", "recursive": true },
      { "path": "./project-skills", "recursive": false }
    ]
  }
}

Implementation Tips

1. Start Small

Don’t load 10 skills at once. Start with one, verify it works, add more gradually.

2. Monitor Performance

Always-on skills consume context. Watch for:

• Slower response times
• Hitting context limits
• Skills conflicting with each other

3. Document Skill Behaviors

When skills are always-on, users may not know they’re active. Document what each skill does:

## Active Skills for Sisyphus Agent

- **content-seed**: Automatically captures interesting patterns, code snippets, and insights during conversation. Seeds are saved to ~/Dev/BrainFucked/seeds/
- **git-workflow**: Enforces issue-driven development. Commits must reference issues.

4. Provide Escape Hatches

Sometimes you need to disable a skill temporarily:

# Start with skill disabled
opencode --disable-skill content-seed

# Or in session
/skill disable content-seed

Key Takeaways

1. Skills as behaviors, not commands - Auto-loading makes skills part of the agent’s personality
2. Per-agent configuration - Different agents need different capabilities
3. Passive > active - The best skills work without being invoked
4. Start small, scale up - Add skills gradually to avoid overload
5. Document what’s active - Users should know what behaviors are running

The shift from manual skill invocation to always-on injection is subtle but powerful. It transforms skills from tools you use into behaviors your agent has.

This pattern emerged from wanting content-seed harvesting to be automatic. The same approach works for any passive skill that should always be active.