Waza
Turns one engineer's proven habits into eight installable agent skills for planning, debugging, review and writing.
- Source repo
- tw93/Waza
- Stars
- ★ 7.3k
- Last updated
- 3d ago
- License
- MIT
- Primary language
- Python
- FA score
- 63/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeChatGPT (Partial support)
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers already using Claude Code or Codex who find model output too generic and want fixed playbooks to raise quality
- Not a fit if
- Locked-down agent environments that forbid third-party skills or plugins
- Teams expecting a large, heavily configurable multi-skill suite
- Users who don't run a host agent like Claude Code, Codex or Cursor
- Source review
- 63/100 · Some gaps
What does this agent do, and when should you use it?
Waza is a small, opinionated collection of agent skills maintained by tw93 that packages engineering habits as slash commands or named skills for Claude Code, Codex, Cursor and similar assistants: /think, /ui, /check, /hunt, /write, /learn, /read and /health. Each skill is a folder containing a SKILL.md playbook, reference docs, helper scripts and gotchas drawn from real failures. Waza ships no model of its own; it runs inside a host agent and is installed with the npx skills CLI into the shared ~/.agents/skills directory, which Claude Code symlinks and other agents read directly. Beyond skills, the repo offers a minimal Claude Code statusline (context window, 5-hour and 7-day quota), a native status_line recipe for Codex, and four optional always-on rules: english, anti-patterns, waza-routing and clarity. It is MIT licensed and positioned by the author as the middle piece of a Kaku / Waza / Kami trilogy, built over 300+ sessions across 7 projects.
Once installed, each skill fires as a slash command or by name inside the host agent. /think challenges the problem and produces a decision-complete plan another agent can implement; /ui builds distinctive frontend interfaces with screenshot-driven aesthetic iteration; /check reviews a diff against project constraints, verifies the result and handles approved release and maintainer actions; /hunt does systematic debugging that confirms root cause before any fix; /write rewrites Chinese and English prose to sound natural; /learn runs a six-phase research workflow (collect, digest, outline, fill in, refine, self-review and publish); /read turns any URL or PDF into a summary or clean Markdown; /health audits Codex, Claude Code, project instructions, verifier output and AI maintainability with a budget-aware summary pass first. Skills chain freely, for example /think then implement then /check, or /hunt then fix then /check. At runtime /check reads the target repository's public context such as READMEs, package manifests, Makefiles and CI workflows, and the README states it never touches private paths, credentials or tokens. Extras include the statusline installer and setup-rule.sh for optional persistent rules.
- Developers already using Claude Code or Codex who find model output too generic and want fixed playbooks to raise quality
- Solo developers or small teams that plan a feature with /think, implement, then run /check before merging
- Anyone fighting a regression that used to work, where /hunt forces root-cause confirmation before a fix
- Frontend builders who want agent-generated UI with a committed direction rather than generic defaults, iterated with /ui
- Researchers and writers who pull sources with /read, synthesize with /learn and polish with /write
- Users auditing their own agent configuration and verify loop with /health
How do you install or deploy this agent?
Prerequisites: Node.js so npx works, and a host agent such as Claude Code, Codex or Cursor. Install the skills into the shared directory with the skills CLI:
npx skills add tw93/Waza -a claude-code codex cursor -g -yThey land in ~/.agents/skills, with Claude Code symlinked in. Update with:
npx skills update -g -yIf you prefer the host's own plugin mechanism (skills are namespaced, e.g. /waza:check), in Claude Code:
/plugin marketplace add tw93/Waza
/plugin install waza@wazaIn Codex:
codex plugin marketplace add tw93/Waza
codex plugin add waza@wazaClaude Desktop users download waza.zip from the latest release and upload it under Customize > Skills; Pi users run pi install npm:@tw93/waza. You can also ask your agent to install Waza by reading llms.txt.
How do you use this agent?
After installation, invoke the matching slash command or skill name in the host agent. Typical chains: settle a feature plan with /think, implement after approval, then /check before merging; ship a fix with /hunt to find root cause, fix, then /check to review and, when needed, handle release, push and issue follow-through; research with /read, synthesize with /learn, polish with /write. Optional always-on rules are installed separately with the script (swap claude-code for codex or antigravity-cli on other hosts):
(
set -e
WAZA_RULE_SCRIPT="$(mktemp -t waza-rule.XXXXXX)"
trap 'rm -f "$WAZA_RULE_SCRIPT"' EXIT
curl -fL https://github.com/tw93/Waza/releases/latest/download/setup-rule.sh -o "$WAZA_RULE_SCRIPT"
# review it first: less "$WAZA_RULE_SCRIPT"
bash "$WAZA_RULE_SCRIPT" english claude-code
bash "$WAZA_RULE_SCRIPT" anti-patterns claude-code
bash "$WAZA_RULE_SCRIPT" waza-routing claude-code
bash "$WAZA_RULE_SCRIPT" clarity claude-code
)For a Codex statusline, add to ~/.codex/config.toml:
[tui]
status_line = ["model-with-reasoning", "current-dir", "context-used", "five-hour-limit", "weekly-limit"]
status_line_use_colors = trueUninstall with npx skills remove tw93/Waza -g and clean up the rule files and statusLine entry as described in the README.
What are this agent's strengths and limitations?
- Deliberately minimal: eight single-purpose skills with explicit triggers, positioned against the author's criticism of heavy, over-configured skill suites
- One npx command installs into ~/.agents/skills, which Claude Code, Codex, Cursor, Gemini CLI, Copilot and others read
- Each skill states the outcome, red lines and verification method rather than scripting every step, so it can ride model improvements
- Gotchas are said to trace to real failures, and /check becomes project-aware by reading the target repo's public context at runtime
- MIT licensed with no dependency on the author's own cloud, plus practical extras like the statusline and four optional persistent rules
- Skills contain no model; quality depends entirely on the host agent's version and capability, with no reproducible benchmarks published
- The 300+ sessions across 7 projects is the author's own account, not a public evaluation
- Documentation has six language variants, but how skill behavior varies by language is not documented
- Installation relies on npx and a third-party skills CLI, while plugin paths differ per host, so cross-tool parity must be checked yourself
- /check and the maintainer actions it handles can push or open issues once approved, so mistakes fall on the user
How does this agent compare with similar options?
The README explicitly contrasts Waza with Superpowers and gstack, calling them powerful but heavy with too many skills and configuration, while Waza opts for eight focused skills. It also sits alongside the author's Kaku (writes code) and Kami (ships documents) as a deliberately different companion.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Waza This agent | 63 · Some gaps | CLIFree | ★ 7.3k | 3d ago | Python | Codex · Claude Code |
| Finding-Unknowns Skills | 76 · Good | Agent plugin / skillFree | ★ 353 | 23d ago | Python | Claude Code · Claude.ai |
| CoCo Super Intelligence | 69 · Some gaps | Agent plugin / skillFree + model costs | ★ 881 | today | Python | Codex · Claude Code |
| Loop Library: Reusable AI Agent Workflows | 61 · Some gaps | Agent plugin / skillFree | ★ 3.2k | 3mo ago | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README states /check reads only public project context (READMEs, manifests, Makefiles, CI) and never private paths, credentials, or tokens; tests assert DO-NOT-PRINT values never appear in output, showing redaction awareness. However the repo is a skill collection whose install scripts curl release assets and pipe to bash (with a 'review first' hint), an external side effect; /check is described as handling release, publish, push, relying on model self-restraint rather than enforced confirmation. Hence least_privilege, user_confirmation, data_flow_transparency, sensitive_data_handling = 2; external_effects = 1 for curl|bash and publish actions; rollback = 2 for documented uninstall; source_attribution = 1 since publisher is unverified and only self-described.
CI pins Python 3.12 and shellcheck with SHA256 verification; release workflow asserts tag matches VERSION and package.json; tests cover MCP config parsing, redaction, path rules and edge cases, so self_consistency and dependency_availability = 2. Failure messages include explicit wait-and-do-not-republish guidance and tests assert exact output strings, so failure_messages = 2; no runtime fallback for model failure documented, so not 3.
README documents install for Claude Code, Codex, Cursor, Gemini CLI, Copilot, Amp, Kimi, Pi, Claude Desktop, so audience_and_scenarios = 2. Eight skills have clear triggers (/think, /ui, /check, /hunt, /write, /learn, /read, /health), capability_boundaries and trigger_precision = 2. environment_fit = 2 for Windows test job and multilingual READMEs, but no guidance for offline or restricted environments.
install_notes is thorough (npx skills, host plugin, Claude Desktop ZIP, Pi, uninstall), = 3. license is full MIT text consistent with package.json, = 3. information_architecture has skill table and structure, = 2. naming_stability = 2 (stable skill names but version 3.40.0 with no changelog file). examples_and_faq = 2 (workflow examples and screenshots). known_limitations = 1 (only scattered notes, no dedicated section). versioning_changelog = 2 (VERSION and release workflow but no CHANGELOG). maintenance_responsibility = 1 (unverified publisher, personal attribution only).
Skills produce actionable artifacts (decision plans, diff reviews, root-cause confirmation), output_usability = 2. marginal_value = 2: contrasts with Superpowers/gstack and emphasizes leanness, but value claims are self-asserted. cost_benefit = 2: eight lightweight skills, one-command install; no quantified benefit so not 3.
claim_traceability = 1: README claims '300+ sessions across 7 projects' without linked evidence. cross_source_corroboration = 2: test files corroborate README claims (redaction, MCP parsing, path rules). fact_inference_separation = 2: README separates facts from design philosophy and tests assert concrete behavior.
- Install scripts curl release assets and pipe to bash; despite a 'review first' hint, prefer manual download and inspection before running.
- /check is described as performing release, publish, and push actions; confirmation relies on model self-restraint, so trial in a controlled repo first.
- Publisher identity is unverified; maintenance responsibility and update path rest on personal attribution only—assess before enterprise use.
- Claims such as '300+ sessions' lack traceable evidence and should not drive adoption decisions.