The best Claude Code debugging skill for most people is Systematic Debugging from the obra superpowers repository: four ordered phases that force the agent to find a root cause before it proposes a fix. Other debugging skills tighten that idea in different ways, with a one-sentence root-cause rule, a reproduce-and-bisect loop, a graph-based tracer or a second opinion from another agent.
This guide compares the root-cause and debugging skills in the debugging topic and the root cause analysis topic. Each entry has a "best for" line, requirements, licence as the directory records it, and a trade-off. The editorial team did not run these skills on real bugs, so nothing here measures how often they find the cause.
Which debugging skill fits which bug?
Match the skill to the kind of failure.
- A test fails or a build breaks and the agent keeps patching symptoms: use a method skill such as obra/systematic-debugging.
- Something that used to work no longer does: tw93/hunt has a bisect mode for it.
- You want a repeatable reproduce, hypothesize, fix loop with a regression test: jsmastery-pro/debug.
- Crashes, hangs, segfaults, CUDA out-of-memory errors or NaN values: stas00/art-of-debugging.
- The agent is stuck in a loop: getpaseo/paseo-committee brings in a second agent.
- CI is red: chachamaru127/ci decides whether the test or the code is wrong.
Best debugging skills compared
| Skill | Best for | Needs | Licence | Automated check |
|---|---|---|---|---|
| obra/systematic-debugging | General bugs and failing tests | Nothing | MIT | Passed |
| tw93/hunt | Evidence-backed diagnosis, regressions | Git | MIT | Passed |
| jsmastery-pro/debug | Reproduce-to-regression-test loop | Git | MIT | Info finding |
| stas00/art-of-debugging | Unix, Python and PyTorch problems | gdb, strace, py-spy as needed | CC-BY-SA-4.0 | Info finding |
| tirth8205/debug-issue | Tracing callers and callees cheaply | The code-review-graph MCP tools | MIT | Passed |
| getpaseo/paseo-committee | Getting unstuck with two agents | Paseo with two agent profiles | No standard licence recorded | Passed |
| chachamaru127/ci | Failing CI pipelines | Git | MIT | Info finding |
"Passed" means the directory's static check reported no findings. An info finding is a low-severity note on the skill's page. Neither says anything about how well a skill debugs.
The reference method: obra/systematic-debugging
Best for: any bug where the agent's first fix did not work.
obra/systematic-debugging lives in the superpowers repository under the MIT licence. Its rule is that no fix comes before a root cause investigation. Phase one reads error messages, reproduces the failure, reviews recent changes and gathers evidence at component boundaries. Phase two compares broken code with a working example. Phase three states a specific hypothesis and tests it with the smallest possible change. Phase four writes a failing test, applies a single fix and checks that nothing else broke.
It also sets a limit: after three different fixes fail, stop proposing more and question the architecture. Three companion files cover backward tracing through call stacks, validation at several layers and waiting on conditions instead of fixed delays.
Requirements: none. At roughly 2,400 tokens it is cheap to keep loaded.
Trade-off: it is a process, not a tool. It will not tell the agent how to read a core dump or profile a GPU. It is also the most copied skill in this comparison, with copies in many other repositories, so install the original rather than a bundle's copy. The superpowers guide explains how it fits with the rest of that skill set.
A separate chriswiles/systematic-debugging in a showcase repository uses the same four phases and the same principle. The file we read names no source, so treat it as an independent variant. The directory records no licence for it, and its last update was earlier this year.
Skills with a different angle
tw93/hunt
Best for: diagnoses that must name one cause and survive review.
tw93/hunt is MIT-licensed. Before touching code the agent must state the root cause in one testable sentence that names a file, function or condition, and the explanation must account for every symptom, not just the first one reported. Words like diagnose and investigate keep the session report-only, while fix and implement unlock edits. It climbs an evidence ladder from reading source to a real runtime check, offers a bisect mode for regressions and a sweep for other instances of the same pattern. After three failed hypotheses it asks for a handoff note listing what was checked and ruled out.
Trade-off: at about 4,300 tokens it is the heavier of the method skills, and the strict gates will slow quick fixes.
jsmastery-pro/debug
Best for: a plain loop that ends with a regression test.
jsmastery-pro/debug is MIT-licensed. Its loop is reproduce, localize, hypothesize, test, fix and verify. It makes the smallest fix, searches for the same root cause elsewhere and asks for exact steps if the bug cannot be reproduced. If the problem points to a design flaw, it hands off to a sibling skill named architect rather than papering over it.
Trade-off: the check raised an informational finding because the skill allows unrestricted shell commands. That is common for debugging skills, but know that the agent can run commands with fewer prompts.
stas00/art-of-debugging
Best for: low-level and machine learning problems.
stas00/art-of-debugging packages a debugging method and tool recipes from an open book. The repository covers Unix tools, compiled programs, Python, PyTorch and machine learning projects. The skill expects tools such as gdb, strace and py-spy, depending on the problem. It is licensed CC-BY-SA-4.0, which has share-alike terms, and the check noted sudo commands in the recipes, so read them before the agent runs anything with elevated rights. At about 6,000 tokens it is the largest here.
tirth8205/debug-issue
tirth8205/debug-issue is a tiny MIT-licensed skill of under 300 tokens. It searches a code knowledge graph, traces callers and callees, follows execution flows, checks recent changes and estimates the impact of a fix, aiming for about five tool calls and a small token budget. It requires the code-review-graph MCP tools for your repository, and its text says the graph supplements reading source and tests, not replaces it.
getpaseo/paseo-committee
getpaseo/paseo-committee forms a committee of two reasoning agents with contrasting profiles, ideally from different model families, that analyze a stuck problem in parallel and reconcile their views. Each prompt ends with an instruction not to edit files, and the skill warns that reasoning can take a quarter of an hour or more. It needs the Paseo tool with two configured profiles, and the directory records no standard licence.
CI, tests and agent runs
chachamaru127/ci diagnoses failing pipelines and tests, decides first whether the test or the implementation is at fault, and hands hard cases to a fixer subagent. Its two extra copies sit in the same repository for other agents, and its check raised the same informational finding about unrestricted shell access.
comet-ml/debugging-e2e-tests shows the project-specific pattern. It investigates a failed Opik end-to-end test, decides regression versus flake and proposes a fix without editing tests, using the Allure TestOps MCP server, the GitHub CLI and Playwright. Skills like this are good templates when you write your own Claude skill.
For failures inside agent sessions, iflytek/orca-replay answers questions about a past run from its recording, using the orcareplay package and its MCP server.
How to choose and install
Start with obra's Systematic Debugging, since it is small and free of requirements. Add hunt when you need a written root cause for review, or the committee skill when the agent loops. Keep project-specific skills for the failures only your codebase has.
Installation steps differ by agent, so use the pages for Claude Code and Codex. Debugging pairs naturally with review, so see the code review skills comparison, and use the top skills list to see what else is widely used.
Frequently asked questions
What does a debugging skill change about how an agent works?
It replaces the agent's habit of guessing a fix with a written procedure: reproduce the failure, find the root cause, test one hypothesis at a time and only then change code. The skill does not give the agent new abilities. It makes the agent slower to edit and quicker to gather evidence, which is usually what a stubborn bug needs.
Which debugging skill is the most widely used?
Systematic Debugging from the obra superpowers repository is the most copied one in the directory, with many other repositories carrying a version of it. It is MIT-licensed and short enough to load on every bug. Other skills here suit narrower jobs such as CI failures or memory leaks.
Do debugging skills work in Codex as well as Claude Code?
Method skills that are plain instructions work in any agent that reads the SKILL.md format. Skills that need an MCP server or a specific command-line tool only work where that tool is available. Check each skill's requirements line, then see the agent pages on this site for where each agent looks for skills.
Can a debugging skill fix a bug by itself?
It can guide an agent to a fix, but you should review the diagnosis and the change. Several skills here stop at a diagnosis or report unless you say fix, and the better ones ask for a failing test before the fix. Treat the agent's root cause as a hypothesis you can verify.
When should I write my own debugging skill?
Write one when the same project-specific failure keeps recurring, such as flaky end-to-end tests or a tricky build setup. The project-specific skills in the directory show the pattern: they name the logs, commands and failure modes unique to one codebase.