safe-debug Skill for Claude: Diagnose ML Failures Before Patching

safe-debug makes your agent diagnose an ML training or inference failure first, and ask before it changes any code.

Installs: 344.7K; on skills.sh
Stars: 486; on GitHub
Files: 3; besides SKILL.md
Updated: Sep 2026; last repo commit

safe-debug triggers when you paste a traceback, terminal error, CUDA out-of-memory message, checkpoint load failure, shape mismatch or NaN loss symptom. It narrows the root cause and writes debug_outputs/DIAGNOSIS.md, PATCH_PLAN.md and status.json.

It does not edit code by default. A patch must be the smallest fix and needs your explicit approval, with a savepoint or branch first for medium or high risk changes. It also says when a fix changes what an experiment means. It reads shared files from ai-research-reproduction, so install the whole set.

Good for

Researchers hitting CUDA OOM, shape mismatches, NaN loss or checkpoint load errors
Engineers who want the smallest fix proposed, not a rewrite
Anyone who must keep debug fixes separate from research changes

Install safe-debug

Claude Code

Copy this prompt.

Install the safe-debug skill in Claude Code. Run this command:

npx skills add lllllllama/RigorPilot-Skills --all

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.

This skill reads shared guideline files from the ai-research-reproduction skill, so install the whole set rather than this folder alone. The README says the main reproduction skill works alone and that "all skills" is the choice for the companion and leaf skills.

From the project README. Source.

What's in the skill

When it loads

Rigor Debug / Rigor Audit skill for deep learning research work. Use when the user pastes a traceback, terminal error, CUDA OOM, checkpoint load failure, shape mismatch, NaN loss symptom, or training failure and wants conservative diagnosis before any patching, with debug fixes clearly separated from research contributions. Do not use for broad refactoring, speculative adaptation, automatic exploratory patching, or general repository familiarization.

Sections in SKILL.md

When to apply
When not to apply
Clear boundaries
Output expectations
Notes

Key facts

Author: lllllllama (owner of the lllllllama/rigorpilot-skills repo)
License: MIT
Installs on skills.sh: 344,707 as of 2026-11-02
GitHub stars: 486 for the whole lllllllama/rigorpilot-skills repo
skills.sh audits: Gen Agent Trust Hub pass, Socket pass, Snyk pass
What is inside: 3 files besides SKILL.md (1 in agents/, 1 in references/, 1 in scripts/)
Needs: The other RigorPilot skills for shared reference files; Python for the bundled helper
Last commit to the repo: 2026-09-23

What it includes

Narrows the root cause from your error
Works from the traceback or symptom you paste before anything is changed.
Try: “Use safe-debug. Here is my traceback from the training run: ...”
Read · Diagnose first
A plan, not an edit
Writes PATCH_PLAN.md and waits for your explicit approval.
Write · Propose the smallest patch
Escalates for a branch or savepoint
For medium or high risk changes it asks for a savepoint or branch first.
Runs code · Savepoint before risky changes
Says when a fix changes comparability
A debug fix is not counted as a research contribution, and SKILL.md has the agent say so when it changes experiment meaning.
Read · Separate debug fixes from research
DIAGNOSIS.md, PATCH_PLAN.md, status.json
The report files named in SKILL.md for each diagnosis. The folder also bundles scripts/safe_debug.py, which SKILL.md does not mention.
Write · Write debug_outputs/

From the official docs

Permissions

Low risk

It diagnoses first and asks before changing code. Read the patch plan before you approve it.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

Patches only happen after you approve
SKILL.md requires explicit approval. Read PATCH_PLAN.md before saying yes.
It bundles a Python script
scripts/safe_debug.py is in the folder, but SKILL.md does not mention it. The README says RigorPilot is local execution, not an OS sandbox. Use the skill on repos you trust.

Technical details

Scopes	Reads your repo and the error you paste Bundles scripts/safe_debug.py; SKILL.md does not say when it runs Writes debug_outputs/ and, after approval, patches code

Questions

What does safe-debug do?

It diagnoses a deep learning failure before anyone patches it. You paste the error, it narrows the cause, and it writes a diagnosis and a minimal patch plan.

How do I install it?

Use the install box at the top of this page, which installs the whole RigorPilot set. The skill reads shared files from ai-research-reproduction, so installing it alone is not enough. The README says the installer needs Node.js/npm and that skills@1.5.26 requires Node 22.20.0 or newer.

Does it need an API key?

No key is named in SKILL.md.

Will it edit my code?

Not by default. SKILL.md says to diagnose first, propose the smallest fix, and require explicit approval before any change.

safe-debug or analyze-project?

Use safe-debug when something is failing now. Use analyze-project for a read-only tour of a repo when nothing is broken.

Is it safe?

It is built to diagnose first: SKILL.md says not to modify repository code by default and to require explicit approval before a patch. Once you approve, it can change your code. The folder also bundles scripts/safe_debug.py, and the README says RigorPilot is local execution, not an OS sandbox, so use repos you trust.

Information

updated Nov 2, 2026

Repository: lllllllama/rigorpilot-skills
Folder: skills/safe-debug

lllllllama/rigorpilot-skills skills.sh

Sources

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

Show all 3

RigorPilot Skills README · github.com
safe-debug SKILL.md · github.com
skills.sh listing and audits · skills.sh