snarktank/amp-skills

Overview

This skill automates feature planning and autonomous execution in two phases: setup and loop. It guides you through creating small, ralph-ready tasks with clear dependencies, saves the parent task context, and then runs a repeatable loop that picks ready tasks, implements them, commits, and moves to the next until the feature is complete.

How this skill works

In Setup mode the assistant asks clarifying questions, breaks the feature into right-sized tasks, creates a parent task and subtasks with dependsOn, and writes the parent-task-id and a fresh progress.txt to scripts/ralph. In Loop mode the agent reads the saved parent-task-id, finds ready leaf tasks (respecting dependsOn), runs each task in a fresh Amp instance, verifies acceptance criteria, commits changes, updates progress, and repeats until all tasks complete.

When to use it

• Starting a new feature and you want automated, iterative implementation
• You already have task_list tasks and want Ralph to run them autonomously
• When tasks must be small, verifiable, and executed sequentially by dependencies
• When you need repeatable progress tracking and a short-term feature memory
• When you want end-to-end implementation from schema to UI to tests

Best practices

• Keep tasks small — each must fit one Ralph iteration and be describable in 2–3 sentences
• Order tasks by dependencies: schema → backend → UI → integration/tests
• Include concrete acceptance criteria (npm run typecheck passes; npm test when applicable)
• Use explicit dependsOn for ordering and avoid circular dependencies
• Preserve Codebase Patterns in progress.txt, archive prior progress before resetting

Example use cases

• Plan and implement an expense upsert tool: create schema migration, backend actions, UI, and tests as separate tasks
• Adopt existing task trees: verify subtasks and dependencies, save parent-task-id, and start the loop
• Split a large dashboard feature into schema, queries, components, filters, and tests, then run Ralph to implement each piece
• Automate repetitive feature builds where each iteration must be isolated and verifiable