jasonraimondi

planning-ralph

Ralph Wiggum AI — autonomous PRD-driven development agent. Use when the user says "ralph", wants to work through a PRD, or wants autonomous feature implementation from a plans/prd.yaml file.

jasonraimondi 20 Updated 3mo ago
GitHub

Install

npx skillscat add jasonraimondi/dotfiles/planning-ralph

Install via the SkillsCat registry.

SKILL.md

Ralph Wiggum AI

Autonomous agent that implements features one at a time from a PRD.

Arguments

  • forever — loop through all remaining tasks until the PRD is complete instead of stopping after one task.

Required Files

  • PRD: ./plans/prd.yaml
  • PROGRESS: ./plans/progress.md
  • RESEARCH: ./plans/research.md

If any required file is missing, stop and tell the user to run the PRD planning skill first.

Mode

  • Default (no arguments): work on a single task, then stop.
  • forever: loop back to step 2 after each successful commit. Stop only when all tasks are done, a task is blocked, or you are unable to commit.

Workflow

  1. Review RESEARCH for relevant context.
  2. Find work: if any task has status: in_progress, continue it. Otherwise, pick the next pending task using this priority:
    • Tasks that unblock other tasks (dependencies first).
    • Phase order (lower phases before higher).
    • Track order: datarulesintentexperienceresilience.
  3. Set the task's status to in_progress in the PRD.
  4. Implement the feature. If the task has tdd: true, follow the RED→GREEN vertical slices in order.
  5. Lint and test: if failures occur, attempt to fix and retry ONCE. If still failing, set status to blocked, document the blocker in PROGRESS, and stop.
  6. UI verification (for tasks with headed: true):
    • Start dev server using dev_command from PRD if base_url is unreachable.
    • Use tooling-agent-browser skill (agent-browser CLI) to navigate, snapshot, and verify each step.
    • Close when done: agent-browser close.
  7. Set the task's status to done in the PRD.
  8. Append to PROGRESS — leave a useful note for the next session.
  9. Stage only the changed files and commit with a concise conventional commit message.
    • If a pre-commit hook fails: fix the issue, re-stage, and retry the commit.
    • If the commit is denied by the user or you are otherwise unable to commit (permission rejected, hook you cannot fix): halt immediately — do not continue to the next task.
  10. If all PRD tasks are done, output COMPLETE and stop.
  11. Forever mode only: if argument is forever and commit succeeded, loop back to step 2.