Metaharness

A Claude Code skill for designing agents, commands, and skills. Harness to write harness.

Install

claude skill install lightfastai/metaharness

What It Does

Metaharness guides you through designing Claude Code harness components with structural consistency and best practices. It handles:

• Agents (.claude/agents/*.md) — isolated subagents for focused subtasks
• Commands (.claude/commands/*.md) — user-invoked slash commands
• Skills (SKILL.md) — domain knowledge + workflow guidance

Usage

/metaharness

Or just describe what you want — the skill auto-triggers on phrases like "create an agent", "design a skill for X", "improve this agent".

What's Inside

metaharness/
├── SKILL.md                     # Main skill — 6-phase workflow
└── resources/
    ├── agent-anatomy.md         # Structural template with all required sections
    ├── best-practices.md        # Design principles + named anti-patterns
    └── evaluation-checklist.md  # Pass/fail quality review criteria

The Workflow

1. Understand Intent — what's the job, what's the boundary, who consumes the output
2. Research Existing — read all agents/commands/skills for consistency
3. Design — apply the anatomy template and best practices
4. Write — produce the component file
5. Evaluate — run the checklist (redundancy, scope, structure, quality)
6. Iterate — refine based on feedback

Best Practices Codified

The skill encodes lessons from production agent design:

• One agent, one job — if you need two sentences to describe it, it's two agents
• Each instruction lives in exactly one section — no redundancy across CRITICAL, Core Responsibilities, Protocol, and What NOT to Do
• Output is a contract — structured templates when agents consume the output, flexible when humans do
• Description is the dispatch trigger — scannable, one sentence, states WHAT not HOW
• Minimum viable tool set — every extra tool is scope creep
• REMEMBER anchor — the last line gets disproportionate weight

Named Anti-Patterns

The evaluation checklist catches these by name:

Anti-Pattern	Symptom
Swiss Army Agent	4+ responsibilities, 8+ tools, does 5 things adequately instead of 1 thing well
Echo Agent	Same instruction repeated in CRITICAL, Responsibilities, Protocol, and What NOT to Do
Rigid Template Agent	Output template can't adapt to query variety
Critic-by-Default Agent	Research agent that spontaneously suggests improvements
Context Hog	Prompt over 150 lines eating into working context

License

MIT