Agent Skills: Terminal-Driven Fix Loop

Use a terminal-first feedback loop where the user runs tools manually and provides output for fixes.

UncategorizedID: diey/codex-agent-skills/terminal-driven-fix-loop

Install this agent skill to your local

pnpm dlx add-skill https://github.com/diey/codex-agent-skills/tree/HEAD/terminal-driven-fix-loop

Skill Files

Browse the full folder contents for terminal-driven-fix-loop.

Download Skill

Loading file tree…

terminal-driven-fix-loop/SKILL.md

Skill Metadata

Name
terminal-driven-fix-loop
Description
Use a terminal-first feedback loop where the user runs tools manually and provides output for fixes.

Terminal-Driven Fix Loop

Goal

Use a tight, terminal-first workflow:

  • The assistant implements changes without running tools
  • The user runs tests/formatters/static analysis manually
  • Fixes are driven by terminal output

Default Workflow

  1. Implement the requested task without running tests, linters, formatters, or static analysis.
  2. End with:
    • What changed (brief)
    • Suggested commands for the user to run (brief)
  3. Wait for terminal feedback.

When the User Pastes Errors

When the user provides terminal output:

  • Use the command, file paths, line numbers, and messages from the output to guide fixes.
  • Apply the smallest fix needed to address the reported error.
  • Do not introduce unrelated refactors while fixing errors.
  • Do not request code snippets unless the error output is ambiguous or lacks enough location/context to act.

What to Ask For (When Needed)

Only request additional context if necessary, for example:

  • The error output does not include a file path or line number
  • The stack trace points to generated/vendor code without showing the caller
  • The failure depends on runtime state that is not visible from output

If asking for more context, ask for the minimum needed.

Tool Execution Policy

  • Never run tests, linters, formatters, or other commands automatically.
  • Do not assume permission to execute tools.
  • You may suggest commands, but do not execute them.

Anti-patterns

  • Running tools without being asked
  • Guessing causes instead of using terminal output
  • Asking for snippets when output already identifies the location
  • Fixing multiple unrelated issues in one pass