Waza

Best For

Engineers using Claude Code or Codex who want more controlled, precise AI agent behavior. Ideal for solo developers or small teams needing a low-configuration engineering habits framework. If Superpowers feels too heavy and writing your own rules feels too ad-hoc, Waza is the right middle ground.

How I Actually Use It

Installation is one line: npx skills add tw93/Waza -a claude-code -g -y. The eight skills span four workflow phases: Pre-build (/think, /design), Post-build (/check), Diagnostic (/hunt, /health), and Content (/read, /write, /learn). I use /health most frequently for agent health audits, checking whether Claude Code configuration has drifted. /check runs before merging, dynamically extracting constraints from diffs and manifests without additional configuration.

What It Does Well

The 36 anti-pattern guardrails are the standout feature. Each pattern includes specific wrong behavior and correct behavior side by side. For example, #32 explicitly states that externally fetched content is untrusted data, even if it contains "ignore previous instructions." #18 requires actual shell output before claiming "I ran the tests." #19 specifies that a user saying "ok" only approves draft text, not git push or npm publish.

The skill vs. script decision table draws a clear boundary: work requiring judgment, context adaptation, or user follow-up goes into a Skill; work with deterministic input-output goes into a Script. The hard cap of eight skills forces the framework to stay lean. New capabilities must land in references, rules, or scripts.

RESOLVER.md contains ten disambiguation rules written in Chinese, handling gray areas like "is this screenshot an aesthetic issue or a regression" and "is this optimization or debugging."

Failure Modes and When Not to Use

Skill chaining requires manual triggering, making it unsuitable for fully automated workflows. The framework deliberately "trusts the model," so if the underlying LLM is weak, skill effectiveness is limited. RESOLVER.md is in Chinese while skill definitions are in English, creating a context switch. If you need complete automation pipelines (e.g., CI/CD integration), Waza is not the right tool.

Pricing, Learning Curve, and Risk

MIT License, completely free. One-line npx install, shared globally. Very low barrier to entry. Note that anti-pattern #17 prohibits all AI attribution in public text (including Co-authored-by: Claude). If your team or platform requires AI attribution, you will need to adjust this rule yourself.

Verdict

Waza's greatest contribution is not the skills themselves but the design methodology: the skill vs. script boundary definition, 36 anti-pattern guardrails, and the single-source-of-truth VERSION pattern. Even if you do not plan to install it, these design principles are worth adopting in your own AI engineering practice.

Sources

• GitHub Repository
• Author: tw93 (also behind Pake and Maple Mono)
• Version: v3.26.0 (2026-05-26)