ccw-cli-tools_skill

Overview

This skill provides a unified CLI execution framework for multi-model code analysis and implementation tasks. It reads a JSON configuration to select tools and models, builds structured prompts with a six-field template, and runs ccw cli commands across analysis, write, and review modes. The framework includes auto-fallbacks, rule templates, and session resume capabilities to support multi-step workflows.

How this skill works

On first use it loads a global configuration (~/.claude/cli-tools.json) or reuses an in-memory copy to discover available tools, models, tags, and types. It maps user intent to a tool and primary model (or fallback), assembles a prompt using PURPOSE/TASK/MODE/CONTEXT/EXPECTED/CONSTRAINTS, and executes ccw cli with the selected rule template and options. Results are returned to the user and failures trigger secondary-model or fallback-tool logic.

When to use it

• Automated code review, security assessments, or architecture analysis in read-only mode
• Feature implementation, refactors, or test generation when write permissions are explicitly requested
• Git-aware reviews using codex for branch diffs, uncommitted changes, or commits
• Multi-round workflows that require session resume and multi-model collaboration
• Scoped inspections that need fine-grained CONTEXT patterns and includeDirs for cross-repo files

Best practices

• Ensure configuration is loaded into memory to avoid repeated file reads
• Always specify MODE explicitly; default to analysis for safe, read-only runs
• Use the 6-field prompt template (PURPOSE, TASK, MODE, CONTEXT, EXPECTED, CONSTRAINTS) with concrete success criteria
• Provide file patterns and --includeDirs when referencing parent or sibling directories
• Attach a specific --rule template for consistent protocol and deliverable structure

Example use cases

• Run a security assessment on the auth module with analysis mode and an assessment rule
• Implement a new middleware feature in write mode with a development-implement-feature rule
• Diagnose a memory leak with an execution-trace or root-cause analysis template
• Perform a codex review of an uncommitted branch or a specific commit using review mode
• Resume a prior planning session to merge findings or continue multi-model work