debuggingclaude-codecursorResearch-backed
Debug
Structured debugging session — reproduce, isolate, diagnose, and fix.
Upstream publishes this file as a Claude skill (SKILL.md). Cursor reads the same SKILL.md shape. This page does not claim either product ships the file.
What the research says
Research-backedObserved professionals reproduce, build a model, fix, and validate, and they lean on recent changes, which is the loop this skill teaches. Offering developers candidate hypotheses made them 6x more likely to succeed. The agent-side gain is unproven: LLM "scientific debugging" did not consistently beat plain repair.
- Li & Coblenz, a grounded theory of debugging (2026)
- Alaboudi & LaToza, hypotheses as a debugging aid
- Kang et al., AutoSD scientific debugging
Checked Oct 17, 2026. How we rate evidence
Let your agent do it
Copy a prompt for Claude Code, Codex, or Cursor.
Paste it into your coding agent in this repo. It does the setup and asks before anything leaves your machine.
Runs in GitHub Actions on your Claude Pro or Max plan through an OAuth token from claude setup-token. No API credits.
Set up the "Read failing CI logs and suggest a fix" GitHub workflow from AgentFlows (https://agentflows.sh/flows/claude-ci-failure-fix) so it runs on my Claude Pro or Max subscription instead of API credits.
Install these agent skills first. Download each SKILL.md from its source (respect its license) into .claude/skills/<name>/SKILL.md, or wherever this agent loads skills from:
- Debug: https://github.com/anthropics/knowledge-work-plugins/blob/main/engineering/skills/debug/SKILL.md (Apache-2.0), notes at https://agentflows.sh/skills/debug
Save this file as .github/workflows/claude-ci-failure-fix.yml:
```yaml
name: CI failure helper
on:
workflow_run:
workflows: ["CI"]
types: [completed]
jobs:
diagnose:
if: github.event.workflow_run.conclusion == 'failure'
runs-on: ubuntu-latest
permissions:
contents: read
actions: read
pull-requests: write
id-token: write
steps:
- uses: actions/checkout@v6
- uses: anthropics/claude-code-action@v1
with:
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
additional_permissions: |
actions: read
prompt: |
REPO: ${{ github.repository }}
RUN ID: ${{ github.event.workflow_run.id }}
Read the failing job logs, find the root cause, and comment a fix on the related PR.
```
It authenticates with CLAUDE_CODE_OAUTH_TOKEN. Tell me to run `claude setup-token` in my own terminal to create it, then run `gh secret set CLAUDE_CODE_OAUTH_TOKEN` and let me paste it. Never write the token into a file.
It runs on: workflow_run: completed (failure). Runs count against my subscription's usage limits, and GitHub Actions minutes still apply.
Show me the diff, then commit on a new branch and open a pull request.The file
SKILL.md
Copy the upstream file, including its own frontmatter, into .claude/skills/debug/SKILL.md.
.claude/skills/debug/SKILL.md
--- name: debug description: Structured debugging session — reproduce, isolate, diagnose, and fix. Trigger with an error message or stack trace, "this works in staging but not prod", "something broke after the deploy", or when behavior diverges from expected and the cause isn't obvious. argument-hint: "<error message or problem description>" --- # /debug > If you see unfamiliar placeholders or need to check which tools are connected, see [CONNECTORS.md](../../CONNECTORS.md). Run a structured debugging session to find and fix issues systematically. ## Usage ``` /debug $ARGUMENTS ``` ## How It Works ``` ┌─────────────────────────────────────────────────────────────────┐ │ DEBUG │ ├─────────────────────────────────────────────────────────────────┤ │ Step 1: REPRODUCE │ │ ✓ Understand the expected vs. actual behavior │ │ ✓ Identify exact reproduction steps │ │ ✓ Determine scope (when did it start? who is affected?) │ │ │ │ Step 2: ISOLATE │ │ ✓ Narrow down the component, service, or code path │ │ ✓ Check recent changes (deploys, config changes, dependencies) │ │ ✓ Review logs and error messages │ │ │ │ Step 3: DIAGNOSE │ │ ✓ Form hypotheses and test them │ │ ✓ Trace the code path │ │ ✓ Identify root cause (not just symptoms) │ │ │ │ Step 4: FIX │ │ ✓ Propose a fix with explanation │ │ ✓ Consider side effects and edge cases │ │ ✓ Suggest tests to prevent regression │ └─────────────────────────────────────────────────────────────────┘ ``` ## What I Need From You Tell me about the problem. Any of these help: - Error message or stack trace - Steps to reproduce - What changed recently - Logs or screenshots - Expected vs. actual behavior ## Output ```markdown ## Debug Report: [Issue Summary] ### Reproduction - **Expected**: [What should happen] - **Actual**: [What happens instead] - **Steps**: [How to reproduce] ### Root Cause [Explanation of why the bug occurs] ### Fix [Code changes or configuration fixes needed] ### Prevention - [Test to add] - [Guard to put in place] ``` ## If Connectors Available If **~~monitoring** is connected: - Pull logs, error rates, and metrics around the time of the issue - Show recent deploys and config changes that may correlate If **~~source control** is connected: - Identify recent commits and PRs that touched affected code paths - Check if the issue correlates with a specific change If **~~project tracker** is connected: - Search for related bug reports or known issues - Create a ticket for the fix once identified ## Tips 1. **Share error messages exactly** — Don't paraphrase. The exact text matters. 2. **Mention what changed** — Recent deploys, dependency updates, and config changes are top suspects. 3. **Include context** — "This works in staging but not prod" or "Only affects large payloads" narrows things fast.