Waza: eight engineering habits packaged as skills for Claude Code, Codex and Cursor

What Waza packages, and why habits instead of prompts

Waza's pitch is modest in a useful way: engineering habits you already know, turned into skills an agent can run. There are eight, and each has a clear moment of use. /think comes before building anything new and produces a plan another agent can implement. /ui is for frontend work and iterates on screenshots toward a committed visual direction. /check reviews a diff against project constraints before merging or releasing. /hunt is systematic debugging that confirms the root cause before any fix. /write rewrites prose to sound natural in Chinese and English. /learn is a six-phase research workflow, /read turns a URL or PDF into a summary or clean Markdown, and /health audits the agent setup itself.

The README says each skill is a folder with reference documents, helper scripts and gotchas from real failures. That last part is the difference from a prompt you paste. A good debugging habit is mostly a list of mistakes not to repeat, and the skills carry that list. The people who benefit are developers who already delegate real work to an agent and have noticed it cutting the same corners.

Installing Waza with npx skills

The README's main install command adds the skills for several agents at once:

npx skills add tw93/Waza -a claude-code codex cursor -g -y

One copy lands in ~/.agents/skills, which the README calls the shared skills directory. Claude Code is symlinked in, and Codex, Cursor, Gemini CLI, Copilot, Amp, Kimi Code CLI and other agents that read that directory pick up the eight skills under the same names. Agents with a private skills directory take their own id after -a. Updating uses the same tool:

npx skills update -g -y

There are other routes. Claude Desktop users upload a waza.zip from the latest release through the Skills settings, and Pi users install the npm package @tw93/waza. The shared-directory approach has one consequence worth knowing: an update changes the skills for every agent that reads the directory at the same moment.

The plugin route for Claude Code and Codex

If you prefer each host's own update command, Waza also ships as a plugin. The skills are then namespaced, so /check becomes /waza:check. In Claude Code:

/plugin marketplace add tw93/Waza
/plugin install waza@waza

And in Codex:

codex plugin marketplace add tw93/Waza
codex plugin add waza@waza

The choice between the two routes is about who manages updates. The npx route gives you one copy and one update command across all agents. The plugin route lets Claude Code and Codex each update on their own schedule, at the cost of namespaced commands and two installs to keep in sync. For one person on one agent, the plugin route is tidier; for anyone switching between agents during the day, the shared directory is the reason to pick npx.

Chaining skills and how /check reads a project

Waza does not force a pipeline. Each skill stops at the outcome you asked for, and the README says an explicitly authorised workflow continues without asking you to approve every transition. It lists the common chains: plan a feature with /think, approve, implement, then /check before merging; ship a fix with /hunt, the fix, then /check; research with /read, /learn and /write; debug and verify with /hunt and /check.

The one skill that adapts to your repository is /check. The README says Waza ships only generic habits, and /check becomes project-aware at runtime by reading the target repository's public context, such as READMEs, package manifests, Makefiles and CI workflows, plus your task constraints, and never private paths, credentials or tokens. That is a sensible boundary for a review skill. It also means /check can only enforce what your repository writes down; a rule that lives only in someone's head will not be checked.

The statusline extra and what to review first

Waza also ships a minimal statusline for Claude Code that shows the context window, the 5-hour quota and the 7-day quota, colour-coded by usage. It installs through a script downloaded from the latest release, and the README's snippet includes a comment telling you to review it with less before running it. Take that advice: it is a shell script from the internet that changes your agent configuration.

Codex needs no script, because it has native statusline items. The README's configuration goes into ~/.codex/config.toml:

[tui]
status_line = ["model-with-reasoning", "current-dir", "context-used", "five-hour-limit", "weekly-limit"]
status_line_use_colors = true

The README points out a difference between the two: Codex shows remaining quota, while the Claude Code statusline shows the used percentage, because upstream does not yet expose used values for the 5-hour and weekly windows. There are also three optional rules installed the same way: English coaching, always-on anti-pattern guardrails and a routing hint for non-Claude hosts.

Releases, tests and the MIT licence

Waza moves quickly. The last three releases, 3.35.0 Ruler, 3.36.0 Intent and 3.36.1 Voice, came out on 2026-09-02, 2026-09-06 and 2026-09-12, and package.json carries version 3.36.1. The Makefile shows how the project guards that pace: its test target verifies the skill documents, checks that generated metadata matches the VERSION file and skill frontmatter, checks routing drift, syntax-checks every shell and Python script it finds, runs Python unit tests when pytest is installed, and runs a smoke test for each tests file.

The licence is MIT, so copying one skill into your own agent configuration and adapting it is allowed. With a release every few days, that may suit teams that want a stable review habit more than the latest wording: take /check or /hunt, pin it, and update deliberately instead of on every npx skills update.

Editorial conclusion

Install Waza if you already work with Claude Code or Codex and want planning, debugging and review to follow the same playbook every time. Start with npx skills add tw93/Waza for the agents you use, try /hunt on the next real bug, and read setup-statusline.sh before running the optional statusline installer.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. tw93/Waza on GitHub