hunt Skill (tw93) for Claude Code: Find Root Cause

Makes Claude state the root cause of a bug in one sentence, backed by evidence, before it touches any code.

Installs: 15.1K; on skills.sh
Stars: 7.5K; on GitHub
Files: 5; besides SKILL.md
Updated: Nov 2026; last repo commit

hunt makes Claude diagnose before it fixes. It must state the cause as one sentence, such as "I believe the root cause is X because Y", naming a file, function or line, and run a probe that would fail if the guess were wrong.

It asks for a reproducible check first and stops after three failed hypotheses with a handoff. Words like diagnose, investigate or why make it report only. It changes code only when you explicitly ask for a fix, and a fix request never covers commit, push or publish. It starts its first line with a 🥷 emoji, as the SKILL.md instructs.

Good for

Developers chasing a crash, regression or flaky failure
People whose agent keeps patching symptoms
Teams that want a regression test with every recurring bug

Install hunt

Claude Code

Copy this prompt.

Install the hunt skill in Claude Code. Run this command:

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

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

The README route installs all eight Waza skills into ~/.agents/skills (one shared copy). hunt works on its own, so the single-skill routes are enough. Pick one route, not several. The plugin routes namespace the skill, for example /waza:hunt. The marketplace file says the single-skill plugin needs Claude Code v2.1.143 or newer.

From the project README. Source.

What's in the skill

When it loads

Finds root cause before any fix. Use when something errors, crashes, regresses, or used to work. Not for code review or new features.

Sections in SKILL.md

Outcome Contract
Diagnosis Signals
Durable Context Preflight
Bisect Mode
Repeated Regression / Screenshot Reference Mode
Scope Blast Mode
Confirm or Discard
Runtime Evidence Ladder
Native App Freeze Mode
Targeted Logging
Rendering Bug Mode
IME / Unicode Issues
Hard Rules
Gotchas
Output
Success Format

Key facts

Author: tw93 (repo owner)
Repo: tw93/waza
Installs on skills.sh: 15,053
Repo stars: 7,488
License: MIT
Last repo commit: 2026-11-01
Files in the skill folder: 6 files including SKILL.md
Part of: Waza, a set of 8 skills by tw93

What it includes

No code until one testable sentence names the cause
The hypothesis must name a specific file, function, line or condition and explain every symptom.
Root cause sentence
Find the commit that broke it
Checks git status first and uses a temporary detached worktree if the checkout is dirty. Reads the diff since the last good version before bisecting.
Runs code · Bisect Mode
Use a good screenshot or version as evidence
Lists every symptom, defines the pass or fail check before editing and names the delta from the reference.
Repeated Regression / Screenshot Reference Mode
Search the repo for the same bug pattern after the fix
Greps for the pattern signature and classes each match as same bug, safe or unsure.
Runs code · Scope Blast Mode
Run the one probe that would disprove the hypothesis
A disproven hypothesis is dropped completely.
Confirm or Discard
Source trace, repro, logs, build or test, real runtime check
Compile-only is not enough for UI, native, rendering or generated-artifact bugs.
Runtime Evidence Ladder
Every log answers a yes or no question
Temporary logs are removed before finishing.
Write · Targeted Logging
Native app freeze, rendering bugs, IME and Unicode
Load references/logging-techniques.md, rendering-debug.md or ime-unicode.md for these cases, and failure-patterns.md for recurring failure classes.
Reference modes
Three failed hypotheses means stop
Also covers performance numbers, magic-number tuning, and diagnosing a failing tool before switching to another.
Hard Rules
Root cause, fix, sibling sweep, confirmation, tests, regression guard
Status is resolved, resolved with caveats, or blocked. A recurring bug needs a regression test that was run red then green.
Output

From the official docs

Permissions

Medium risk

The skill runs diagnostic commands in your project and, when you ask for a fix, edits files. Keep your work committed first, and check the diff before you keep any change.

What it can do

Read your data
Create and change things
Delete things
Read-only modeHow to turn it on

Watch out for

It runs commands and may edit code
To reproduce a bug it runs repro commands, tests and builds, and after you ask for a fix it edits files. Review the diff, and keep unrelated work committed or stashed. github.com
Bisect uses git and a temporary worktree
The SKILL.md says a dirty checkout means bisecting in a temporary detached worktree that it removes afterwards, and that git bisect reset runs before removal. github.com
A fix request does not cover commit or push
The SKILL.md says a fix authorization never carries commit, push, publish or any destructive action with it. Ask for those separately. github.com

Technical details

Turning on read-only	Diagnose, investigate, why, look into and similar words make it report only. A fix needs an explicit request.

Questions

What does the hunt skill do?

It makes Claude find the root cause of an error, crash or regression before it fixes anything. The cause has to be one specific, testable sentence backed by evidence.

Will it fix my bug or only explain it?

Words like diagnose, investigate and why get a report only. It edits code when you ask to fix, change or implement, and even then it must prove the cause first.

What if it cannot find the cause?

After three failed hypotheses it stops and writes a handoff: the symptom, each hypothesis and why it was ruled out, the evidence, what is unknown and what it needs from you.

How is it different from the think skill?

hunt is for errors, crashes and regressions. think plans new features and architecture, and routes bug reports to /hunt.

Why does Claude start its reply with a ninja emoji?

The SKILL.md tells it to prefix the first line with 🥷 inline. It is part of the Waza skill set.

How do I install it?

Use the install box above. The Waza README documents the skills CLI for several agents and a plugin route. The skill folder has 6 files including SKILL.md.

Is it safe to install?

It bundles no scripts, but it directs Claude to run commands and edit code in your project. It only edits after an explicit fix request, so keep your work committed and review the diff.

Information

updated Nov 2, 2026

Repository: tw93/waza
Folder: skills/hunt

tw93/waza skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 7

Waza README · github.com
Waza marketplace.json · github.com
hunt SKILL.md · github.com
hunt references/failure-patterns.md · github.com
hunt references/logging-techniques.md · github.com
hunt references/durable-context.md · github.com
Waza README · github.com