Commands vs Agents vs Skills in Claude Code: When to Use Each
A decision framework for choosing between Commands, Sub-Agents, and Skills in Claude Code. Architecture, trade-offs, and real examples of each pattern.
AI for the 1%.
Advanced AI, no filler.
This site is for developers, automation builders, and power users who already use AI daily and want to push harder.
Real configs from a production setup. Honest opinions on what works. Deep-dives that skip the basics.
No ads. No affiliate links. No sponsored content. No beginner hand-holding. If you need that, AiForThe99.com is built for it.
Browse the deep-dives
Every article assumes you know the basics and goes straight to implementation. Real prompts, real configs, real results.
Read the deep-divesExplore by subject
Claude Code, agent workflows, automation patterns, tool reviews. Find what matches your current project.
Browse by subject
Who this site is for
Developers
Ship faster with AI-assisted code, automated reviews, and prompts that understand your codebase. Claude Code, custom skills, CI/CD patterns.
Automation builders
Build AI teams, not just use AI tools. Multi-agent orchestration, pipeline design, workflow systems that improve themselves over time.
Power users
Pick the right model for the right job. Tool deep-dives, homelab setups, prompt patterns, and the decision frameworks behind every choice.
How this site works
1
We build and run AI systems daily
Multi-agent pipelines, custom skills, automated publishing, homelab infrastructure. This is our production setup, not a demo.
2
We write about what we learn
Every article comes from real experience. What worked, what didn't, and the actual configs behind it.
3
You take what's useful
Copy the prompts, adapt the configs, steal the patterns. Everything is designed to be used, not just read.
4
Community (coming soon)
We're building a Skool community for sharing configs, prompt patterns, and honest tool discussions with other power users.
Latest Deep-Dives
Claude Code Best Practices in 2026: What Actually Works
Opinionated guide to Claude Code workflows that hold up in production. CLAUDE.md management, cross-model dispatch, context control, and the tips that actually matter.
A practical guide to reading AI benchmarks, understanding what they measure, and using real usage data to choose the right model.
How to design AI workflows with built-in feedback loops that get better every cycle. Real architecture from a working content pipeline.
A deep dive into TEN 07-25, the Department of Labor's AI literacy framework. Five content areas, seven delivery principles, and what they mean for anyone building AI training programs.
A two-agent code review pattern where the implementer and reviewer operate in separate contexts. Architecture, configs, and real bugs caught.
Most CLAUDE.md files are a dumping ground. Here's how to build a three-tier memory system that gives your AI agent exactly the context it needs.
Build persistent, scheduled AI workflows that run without you. Covers role mapping, manual-first automation, scheduled tasks, and real examples from production use.
Put your AI agent and your knowledge base in the same folder. Here's the setup that turns Obsidian into a shared workspace your AI can actually use.
Practical guide to running multiple AI agents that coordinate, remember, and operate on a schedule. Real configs, real costs, real failure modes.
Build a custom Claude skill from scratch. Covers folder structure, SKILL.md anatomy, description optimization, testing, and iteration.
A practical framework for choosing the right AI model for each task. Covers OpenRouter for comparison testing, model categories, cost analysis, and cross-model workflows.
Quick Takes
View allClaude Code: --dangerously-skip-permissions
If you’re running CC in a sandboxed environment or CI pipeline where you trust everything, --dangerously-skip-permissions removes the approval prompts entirely. Massive time saver for automated workflows.
Don’t use it on your main machine unless you enjoy living dangerously. The name is the warning.
CLAUDE.md > System Prompts
If you’re still writing long system prompts for every conversation, try this instead: put your persistent context in a CLAUDE.md file at your project root. Claude Code reads it automatically. No copy-pasting, no token waste on repeated instructions.
Keep it under 200 lines. Point to deeper docs instead of inlining everything. Think of it as working memory, not a manual.
Skills vs Agents vs Commands
Quick mental model:
Skill = a recipe. Static instructions the AI reads before doing work. No runtime, no state. Agent = a worker. Runs autonomously, has tools, makes decisions. Command = a trigger. Shortcut that loads a skill or kicks off a workflow.
Most people over-engineer agents when a skill would do. Start with skills. Graduate to agents when you need persistence or tool use.
The Spec-First Workflow
Stop asking AI to “build X.” Instead: write a spec (even 10 lines), then hand it to Claude Code or Codex.
The spec forces you to think about what you actually want before the AI starts coding. Result: fewer rewrites, better architecture, and a document you can hand off to anyone.
Specs are leverage. The 15 minutes you spend writing one saves hours of back-and-forth.
Hot Take: Most AI Wrappers Are Unnecessary
If your “AI tool” is just a form that sends a prompt to the API and returns the response, you don’t have a product. You have a UI for something the user could do in 30 seconds with a chat window.
Real value comes from context injection, workflow automation, persistent memory, or domain-specific tooling. If your app doesn’t do at least one of those, reconsider.