ralph-executor

name	ralph-executor
description	Detailed Ralph execution guidance. Load when running the Ralph autonomous loop to get step-by-step implementation protocol.

Ralph Executor Skill

Supplements the Ralph rule with detailed execution protocol. Load this skill when running /ralph-loop.

Pre-Execution Checklist

Before starting any task, verify:

• AGENTS.md exists and contains build/test commands
• prd.json exists with at least one pending task
• progress.log exists — if missing, create it:

  # Ralph Wiggum Progress Log
  # Started: [YYYY-MM-DD HH:MM:SS]
  

Implementing a Task

1. Read the task's acceptance_criteria — these are your binary exit conditions
2. Verify all dependencies have status == "complete" in prd.json
3. Read all relevant existing files before making any edits
4. Apply changes incrementally — one logical change at a time
5. Run the verification command from AGENTS.md after each significant change

prd.json Status Transitions

When starting a task:

{ "id": "task-XXX", "status": "in-progress" }

When completing a task:

{ "id": "task-XXX", "status": "complete" }

Only transition to complete after verification passes.

Verification Protocol

Run the test/typecheck/lint command from AGENTS.md.

If it passes:

• Update prd.json → "complete"
• Append completion entry to progress.log
• Create git commit

If it fails:

1. Read the full error output — do not skim
2. Identify root cause, not just the error message location
3. Apply one targeted fix
4. Re-run verification
5. If still failing: append [ERROR] to progress.log, halt, report to user

Writing to progress.log

Always append (never overwrite). Use exact format:

[2026-01-30 14:22:05] Started task-001: Implement user model
[2026-01-30 14:35:12] Completed task-001: User entity with bcrypt hashing, all tests pass
[2026-01-30 14:35:30] Git commit: feat(auth): implement user model with password hashing
[2026-01-30 14:36:00] [ERROR] task-002: TypeScript type mismatch in auth middleware — halted

Quality Rules

From Arize Research (10-15% SWE-bench improvement):

• Never modify code without first reading and understanding it
• Never apply a workaround that hides an error from the compiler or test runner
• Do not delete or weaken existing tests to make CI pass
• Backwards compatibility is assumed unless the task explicitly states otherwise
• Commit messages must explain WHAT changed, not HOW it was changed
• Error messages must remain technically accurate — never swallow exceptions silently