# Diagnose Deep Learning Failures Safely

Deep learning failures can invite risky, premature changes. This skill classifies concrete errors and prepares conservative guidance before any code modification.

## Install

```bash
npx skillstore add lllllllama/safe-debug
```

## Metadata

- Status: approved
- Slug: lllllllama-safe-debug
- Skillstore revision: r2
- Version status: missing
- Tree hash: 9fba62ecf9a8aa3c06cebc01a7294ebbd4a3067714ee3d34ecff624db508f194
- Author: lllllllama
- GitHub username: lllllllama
- License: MIT
- Repository: https://github.com/lllllllama/rigorpilot-skills/tree/main/skills/safe-debug/
- Ref: 656be3040aef5c047555a908cd5c695d22a4a548
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, filesystem
- Quality score: 80
- Quality tier: silver
- Public page: https://skillstore.pages.dev/skills/lllllllama-safe-debug
- Manifest: https://skillstore.pages.dev/api/skills/lllllllama-safe-debug/manifest

## Capabilities

- Classifies supplied errors into CUDA, checkpoint, distributed, device, shape, loss, file, runtime, or unknown categories.
- Accepts inline error text or reads an error from a UTF-8 text file.
- Recommends category-specific checks while withholding patch authorization.
- Identifies cases where a savepoint is recommended before patching.
- Generates diagnosis, patch plan, and status reports in a selected output directory.
- Can emit the analysis to standard output for integration with other tools.

## Use Cases

- Triage a Training Failure: Classify a traceback and identify low-risk checks before changing model code.
- Review a Checkpoint Error: Separate model variant, checkpoint source, and loading strictness issues before adding compatibility logic.
- Protect Experiment Comparability: Flag fixes that could change experiment meaning and require an approval-gated patch plan.

## Prompt Templates

### Diagnose a CUDA Error

```
Diagnose this CUDA out-of-memory error: {{error}}. Identify the likely cause and suggest configuration-only checks. Do not modify code.
```

### Analyze a Checkpoint Failure

```
Review this checkpoint loading failure: {{traceback}}. Check model variant, checkpoint source, and strictness assumptions. Propose the smallest fix without patching.
```

### Narrow a Distributed Failure

```
Analyze this distributed training failure: {{error}}. Use this launch command: {{command}}. Separate environment, configuration, and code hypotheses. Recommend a savepoint if needed.
```

### Audit Research Impact

```
Audit this training failure: {{traceback}}. Consider recent change {{change}} and experiment goal {{goal}}. Explain evidence, uncertainty, reproducibility impact, and a minimal approval-gated plan.
```

## Limitations

- Classification uses keyword rules and does not inspect repository code.
- The skill does not apply patches or verify that a proposed fix works.
- The diagnosis report includes only the first twelve lines of supplied error text.
- Shared policy references outside the skill directory may be unavailable after installation.

## Best Practices

- Provide the complete traceback, launch command, and recent relevant change.
- Review the diagnosis and approve a specific minimal fix before editing code.
- Create a savepoint when the proposed change could affect experiment meaning.

## Anti Patterns

- Do not use this skill for broad repository exploration or refactoring.
- Do not treat keyword classification as proof of the root cause.
- Do not apply suggested changes without reviewing research comparability.

## Security Audit

- Audited at: 2026-08-08T09:07:20.089\+00:00
- Summary: All external-command findings are false positives caused by Markdown backticks in Python strings or documentation; no commands are executed. Both traversal findings are fixed references to shared Markdown guidance, not user-controlled paths. No prompt injection, exfiltration intent, or other semantic threat was found.

## Stats

- Views: 3
- Downloads: 7
- Favorites: 0
- Popularity score: 0