Claude Code workflows

Reproducible practice

Diagnose and fix one failing test without scope drift

Separate reproduction, localization, implementation, and regression review so a plausible patch is not confused with a verified fix.

Prerequisites

• A reproducible failing command and the first actionable error.
• A known baseline git state and permission to edit the affected area.
• A rollback path such as a task branch or reviewed working-tree diff.

Steps

1. Reproduce before prompting

   Capture the exact command and smallest stable error excerpt; redact secrets and personal data.

   # Run the repository-specific failing test
   git status --short

2. Localize in plan mode

   Ask for the failure path, candidate files, competing hypotheses, and a narrow validation plan.

   claude --permission-mode plan

3. Implement one hypothesis

   Allow edits only after the plan identifies why the observed failure should change. Avoid opportunistic refactors.

4. Run the evidence ladder

   Run the original failing check first, then adjacent tests, then the broader project gate.

   # 1. Original failing command
   # 2. Adjacent test group
   # 3. Project lint, typecheck, or build
   git diff --check

Expected result

The original failure is no longer reproducible, adjacent behavior remains intact, and the diff contains only evidence-supported changes.

Validation

• The original command must pass, not merely a newly added test.
• Review the diff for unrelated formatting, dependency, and generated-file changes.
• Document any broader gate that could not run and why.

Failure handling

Cleanup or rollback

• Use the reviewed diff or task branch to remove only the failed experiment.
• Restore generated files with their owning generator, not by hand.

Boundaries and when not to use it

• This workflow is not enough for destructive migrations, incident response, or changes requiring external production verification.
• A passing local test does not prove deployment, permissions, network, or data-dependent behavior.