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.
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 aredone, a task isblocked, or you are unable to commit.
Workflow
- Review RESEARCH for relevant context.
- Find work: if any task has
status: in_progress, continue it. Otherwise, pick the nextpendingtask using this priority:- Tasks that unblock other tasks (dependencies first).
- Phase order (lower phases before higher).
- Track order:
data→rules→intent→experience→resilience.
- Set the task's status to
in_progressin the PRD. - Implement the feature. If the task has
tdd: true, follow the RED→GREEN vertical slices in order. - 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. - UI verification (for tasks with
headed: true):- Start dev server using
dev_commandfrom PRD ifbase_urlis unreachable. - Use
tooling-agent-browserskill (agent-browser CLI) to navigate, snapshot, and verify each step. - Close when done:
agent-browser close.
- Start dev server using
- Set the task's status to
donein the PRD. - Append to PROGRESS — leave a useful note for the next session.
- 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.
- If all PRD tasks are
done, output COMPLETE and stop. - Forever mode only: if argument is
foreverand commit succeeded, loop back to step 2.