Andisab Casdk Harness Cgf Optimize

CGF Optimize Skill

This skill launches the CGF (Claude Gradient Feedback) optimization pipeline for a specified resource or creates a new resource from description.

Usage

Optimization Mode (Existing Resource)

/cgf-optimize <resource> <optimization_goal> [--review]

Creation Mode (New Resource)

/cgf-create <description> [--review]

Arguments

For optimization mode:

• resource: Resource identifier - can be:

  • Agent name: python-expert, refactor-agent
  • Namespaced agent: research-team:research-specialist
  • Full path: .claude/agents/dev-python-expert.md

• optimization_goal: What to optimize for:

  • async programming
  • better error handling
  • code quality improvements
  • Context7 usage patterns

For creation mode:

• description: Natural language description of the desired resource:

  • Python async expert that helps with asyncio patterns
  • Kubernetes deployment agent for managing k8s resources
  • Code review skill for security-focused reviews

• --review (optional): Enable checkpoint mode for human review at each phase

Examples

Basic Optimization

/cgf-optimize python-expert async programming

Runs full optimization pipeline automatically.

With Review Checkpoints

/cgf-optimize typescript-expert --review

Pauses after research, test generation, and evaluation for your review.

Plugin Agent

/cgf-optimize research-team:research-specialist Context7 integration

Optimizes a plugin agent.

Create New Agent

/cgf-create Python async expert that helps with asyncio patterns

Creates initial agent draft using context-engineer, then optimizes.

Create With Review

/cgf-create Kubernetes deployment agent --review

Creates and optimizes with human review at each phase.

Workflow

Optimization Mode

1. INIT: Creates workspace, detects resource type
2. RESEARCH: Investigates domain best practices (via research-team)
3. RESEARCH_ITERATE: Agentic optimization using research findings and LLM self-critique
4. EVALUATE: Assesses results, recommends accept/refine/reject
5. FINALIZE: Applies recommendation

Creation Mode

1. INIT: Creates workspace, detects creation mode
2. CREATE: Spawns context-engineer to create initial resource draft
3. RESEARCH: Investigates domain best practices
4. RESEARCH_ITERATE: Agentic optimization using research findings and LLM self-critique
5. EVALUATE: Assesses results, recommends accept/refine/reject
6. FINALIZE: Applies recommendation

Output

Results saved to workspace/{resource_id}/:

• run_state.json - Current state (supports resume)
• {resource_id}-v{N}.md - Optimized version
• reviews/v{N}_review.md - Evaluation report

Resume

If optimization was interrupted, simply re-run the same command - it will resume from the last checkpoint.

Source: andisab/casdk-harness — distributed by TomeVault.