Agent Skills: Migrating Agent Systems

Migrating Agent Systems

Overview

Migrating agent systems IS orchestrating component creation through skills.

Don't create files directly—invoke the appropriate skill for each component type.

Core principle: Skills encode best practices. Direct file creation bypasses them.

Violating the letter of the rules is violating the spirit of the rules.

Task Initialization (MANDATORY)

Before ANY action, create task list using TaskCreate:

TaskCreate for EACH task below:
- Subject: "[migrating-agent-systems] Task N: <action>"
- ActiveForm: "<doing action>"

Tasks:

1. Project discovery
2. Architecture planning
3. Create constitution (CLAUDE.md)
4. Create rules
5. Create hooks
6. Create skills (if needed)
7. Validate system

Announce: "Created 7 tasks. Starting execution..."

Execution rules:

1. TaskUpdate status="in_progress" BEFORE starting each task
2. TaskUpdate status="completed" ONLY after verification passes
3. If task fails → stay in_progress, diagnose, retry
4. NEVER skip to next task until current is completed
5. At end, TaskList to confirm all completed

Task 1: Project Discovery

Goal: Understand the project before configuring agent system.

Invoke project-discovery skill (if available) or manually:

Discover:

• Tech stack (language, framework, build tools)
• Project structure (key directories)
• Existing workflows (CI/CD, testing, linting)
• Conventions (naming, architecture patterns)

Output: Discovery report with key findings.

Verification: Have documented tech stack and structure.

Task 2: Architecture Planning

Goal: Design the agent system components.

Invoke agent-architect skill (if available) or manually plan:

Component Planning

| Component | Purpose | Needed? | |-----------|---------|---------| | CLAUDE.md | Constitution with laws | Always | | Rules | Path-scoped conventions | If conventions exist | | Hooks | Quality gates | If linting/formatting needed | | Skills | Project-specific capabilities | If repeated workflows | | Agents | Specialized subagents | If isolated tasks needed |

Output: List of components to create with details.

Verification: Have component list with rationale for each.

Task 3: Create Constitution (CLAUDE.md)

Goal: Create the project's CLAUDE.md with appropriate laws.

CRITICAL: Invoke the writing-claude-md skill.

Do not write CLAUDE.md directly. The skill ensures:

• Proper <law> block format
• Verifiable, actionable laws
• Self-Reinforcing Display law
• Quality review via claudemd-reviewer

Typical laws to consider:

• Communication style (language, verbosity)
• Code conventions (if critical)
• Version sync requirements
• Documentation requirements

Verification: CLAUDE.md created and passes claudemd-reviewer.

Task 4: Create Rules

Goal: Create rule files for project conventions.

For each convention identified in discovery:

CRITICAL: Invoke the writing-rules skill.

Common rules:

• Code style (paths: **/*.ts)
• API conventions (paths: src/api/**)
• Test conventions (paths: **/*.test.ts)

Skip if: No conventions need enforcement.

Verification: All planned rules created.

Task 5: Create Hooks

Goal: Create quality gate hooks.

For each quality check needed:

CRITICAL: Invoke the writing-hooks skill.

Common hooks:

• Linting (ESLint, Ruff)
• Formatting (Prettier)
• Type checking (TypeScript, mypy)

Skip if: No quality gates needed.

Verification: All planned hooks created and working.

Task 6: Create Skills (if needed)

Goal: Create project-specific skills.

Only create skills for:

• Repeated workflows specific to this project
• Complex processes that need documentation
• Team knowledge that should be encoded

CRITICAL: Invoke the writing-skills skill.

Skip if: No project-specific capabilities needed.

Verification: All planned skills created.

Task 7: Validate System

Goal: Verify the entire agent system works.

Checklist:

• [ ] CLAUDE.md exists with <law> block
• [ ] Laws display at start of responses
• [ ] Rules inject on matching files
• [ ] Hooks block violations
• [ ] Skills activate on triggers

Test sequence:

1. Start new Claude Code session
2. Verify laws display
3. Work on files matching rules—verify injection
4. Write violating code—verify hook blocks

Verification: All components work together.

Skill Invocation Reference

| Component | Skill to Invoke | |-----------|-----------------| | CLAUDE.md | writing-claude-md | | Rules | writing-rules | | Hooks | writing-hooks | | Skills | writing-skills | | Agents | writing-subagents |

NEVER create these files directly. Always use the skill.

Red Flags - STOP

These thoughts mean you're rationalizing. STOP and reconsider:

• "I can write CLAUDE.md directly"
• "Skip discovery, I know this stack"
• "Don't need rules, just laws"
• "Hooks are overkill"
• "Skip validation, I used the skills"

All of these mean: You're about to create a weak agent system. Follow the process.

Common Rationalizations

| Excuse | Reality | |--------|---------| | "Write directly" | Skills encode best practices. Use them. | | "Skip discovery" | Assumptions lead to misfit systems. Discover first. | | "Just laws" | Laws + rules + hooks = defense in depth. | | "Hooks overkill" | Hooks catch what humans miss. Worth it. | | "Skip validation" | Untested systems fail in production. Validate. |

Flowchart: Agent System Migration

digraph migrate_agent {
    rankdir=TB;

    start [label="Setup agent\nsystem", shape=doublecircle];
    discover [label="Task 1: Project\ndiscovery", shape=box];
    plan [label="Task 2: Architecture\nplanning", shape=box];
    claude_md [label="Task 3: Create\nCLAUDE.md", shape=box];
    rules [label="Task 4: Create\nrules", shape=box];
    hooks [label="Task 5: Create\nhooks", shape=box];
    skills [label="Task 6: Create\nskills", shape=box];
    validate [label="Task 7: Validate\nsystem", shape=box];
    valid [label="All\nworking?", shape=diamond];
    done [label="System ready", shape=doublecircle];

    start -> discover;
    discover -> plan;
    plan -> claude_md;
    claude_md -> rules;
    rules -> hooks;
    hooks -> skills;
    skills -> validate;
    validate -> valid;
    valid -> done [label="yes"];
    valid -> plan [label="no\nrevisit"];
}

References

• project-discovery skill for analyzing projects
• agent-architect skill for planning components
• All writing-* skills for creating components