Two types of skills, three permission modes, and a settings allowlist -- learn how to configure Claude Code for both safety and speed.
Most people think of skills as prompts. That is only half the picture.
Type 1: Capability uplift skills
These improve what Claude does by default. A frontend design skill might instruct Claude to always follow a specific design system and produce animations. A code review skill might add deeper security checks.
These are valuable -- but they may become less necessary as base models improve. The models get better at design and code review over time.
Type 2: Encoded preference skills
These encode your specific workflow. Not what Claude should do in general, but the exact sequence of steps you follow for a task.
Example: Your process for writing a client proposal might be: research the client industry, pull their specific pain points from discovery notes, write the problem framing first, draft the solution description, then add the ROI section with their specific numbers.
That is your workflow. An encoded preference skill captures it so Claude follows your exact process every time -- regardless of what the base model would do by default.
Encoded preference skills are durable. They encode your process, not AI capability. They get more valuable over time.
Anthropic added a skill-creator skill that helps you build new skills from scratch.
Invoke it with /skill-creator and it will guide you through:
- The name and purpose of this skill
- The sequential workflow this skill should follow
- The inputs and expected outputs
It generates the skill markdown file and saves it to your ~/.claude/skills/ folder.
Evaluating skill quality:
Run the same prompt 5-10 times with the skill active. Measure whether the output is consistent. If you get dramatically different outputs each time, the skill steps are too vague. Tighten the sequential steps until consistency improves.
Mode 1: Ask before edits (default)
Claude asks for permission before any action that modifies the system. Safe. Slower. Good for exploration and unfamiliar projects.
Mode 2: Bypass permissions
Claude does anything without asking. No guardrails. The problem: Claude does not classify actions as risky or safe. It bypasses all checks equally. A misunderstood instruction can delete files, push broken code, or post to production without warning.
Mode 3: Auto mode (new -- replaces bypass)
Claude classifies each tool call before executing it. Safe actions proceed automatically. Risky actions still ask for confirmation.
This gives you the speed of bypass permissions without the danger. Claude does the risk assessment -- it does not skip it.
How to enable auto mode:
- CLI: claude --enable-auto-mode
- VS Code: Settings > Claude Code > Auto Mode
- Note: currently in research preview for Claude Team plan
For production use cases, neither bypass nor auto mode may be right. What you want is a precise allowlist.
In your Claude Code settings, define:
- Allowed tools: specific tools Claude can use without asking
- Denied tools: tools Claude can never use
Example: allow file reads and code edits, deny any git push, deny any file deletion.
This is the most reliable approach for workflows where you know exactly what Claude should and should not be able to do.
Config location: ~/.claude/settings.json (user scope, applies globally)
Next lesson: Memory 2.0 -- how Claude Code now consolidates memory automatically while you sleep.
HivePowered AI — AI Like a Pro Training