// the find
tw93/Waza
🥷 Engineering habits you already know, turned into skills Claude can run.
Waza is a set of eight slash-command skills (`/think`, `/ui`, `/check`, `/hunt`, `/write`, `/learn`, `/read`, `/health`) that bundle markdown playbooks, reference docs, and helper scripts for Claude Code, Codex, Cursor, and other agents that read a shared skills directory. It's aimed at developers who want their AI agent to follow a repeatable process for things like debugging, pre-merge review, and UI iteration instead of winging it each time.
It actually has a test suite — pytest for the Python helper scripts and shell tests for the installers — which is rare for a prompt/skills repo where most projects ship zero verification. The scope is deliberately narrow: eight skills, each with one job, explicitly positioned against heavier competitors like Superpowers for being easier to actually use. It's agent-agnostic by design, installing once into `~/.agents/skills` so Codex, Cursor, Gemini CLI, and Copilot pick up the same skills Claude Code uses, rather than being a single-vendor lock-in. Each skill folder keeps reference docs and documented failure cases (`durable-context.md`, `failure-patterns.md`) alongside the skill itself, so the gotchas are versioned instead of living in someone's head.
There's no way to measure whether any of this actually improves agent output — a skill is a markdown file telling the model what to do, and compliance depends entirely on the host agent choosing to read and follow it, with no enforcement if it doesn't. The 'Python' language tag is misleading: the bulk of the repo is markdown instructions and shell/Python support tooling (checks, packaging, tests), not an application you'd import or run. Several install and update paths rely on `curl | bash` (statusline, optional rules) — they tell you to review the script first, but that's still a supply-chain trust pattern worth being deliberate about before piping into a shell. Behavior will drift across hosts since Claude Code, Codex, and Cursor interpret the same SKILL.md differently, and the repo has no way to verify a given host is actually honoring a skill's red lines.