Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Generates structured .code-task.md files from descriptions or PDD implementation plans. Auto-detects input type, creates properly formatted tasks with Given-When-Then acceptance criteria.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 133% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -4% | 0% |
Generate structured code task files from rough descriptions or PDD implementation plans. Auto-detects input type and creates properly formatted .code-task.md files. For PDD plans, processes one step at a time to allow learning between steps.
These rules apply across ALL steps:
.ralph/tasks/): Output directory for code task filesConstraints:
Check if input is a file with PDD plan structure (checklist + numbered steps). Set mode to "pdd" or "description" and inform the user.
For both modes: create measurable acceptance criteria in Given-When-Then format and prepare a task breakdown plan.
Present the proposed breakdown to the user:
Create files following the Code Task Format below.
PDD mode specifics:
.ralph/tasks/{task_name}/step{NN}/ folder by default (zero-padded: step01, step02, step10)task-01-{title}.code-task.md, task-02-{title}.code-task.mdAll tasks:
status: pending, created: YYYY-MM-DD, started: null, completed: null.code-task.md extensionList generated files with paths. For PDD mode, include the step's demo requirements. Suggest running code-assist on tasks in sequence, or using Ralph for autonomous implementation.
Ask: "Would you like me to set up Ralph to implement these tasks autonomously?"
If yes, create a concise PROMPT.md with objective, spec directory reference, execution order, and acceptance criteria. Suggest the appropriate command:
ralph run --config presets/pdd-to-code-assist.ymlralph run -c ralph.yml -H builtin:code-assistEach code task file MUST follow this structure:
markdown--- status: pending created: YYYY-MM-DD started: null completed: null --- # Task: [Task Name] ## Description [What needs to be implemented and why] ## Background [Context needed to understand the task] ## Reference Documentation **Required:** - Design: .ralph/specs/{task_name}/design.md **Additional References (if relevant to this task):** - [Specific research document or section] **Note:** Read the design document before beginning implementation. ## Technical Requirements 1. [First requirement] 2. [Second requirement] ## Dependencies - [Dependency with details] ## Implementation Approach 1. [Implementation step or approach] ## Acceptance Criteria 1. **[Criterion Name]** - Given [precondition] - When [action] - Then [expected result] ## Metadata - **Complexity**: [Low/Medium/High] - **Labels**: [Comma-separated labels] - **Required Skills**: [Skills needed]
Description mode input: "I need a function that validates email addresses and returns detailed error messages"
Description mode output: .ralph/tasks/email-validator.code-task.md — task with acceptance criteria for valid/invalid email handling, error messages, and unit tests.
PDD mode input: ".ralph/specs/data-pipeline/plan.md"
PDD mode output: .ralph/tasks/data-pipeline/step02/ containing task-01-create-data-models.code-task.md, task-02-implement-validation.code-task.md, task-03-add-serialization.code-task.md — each with design.md reference, acceptance criteria, and demo requirements.
Vague description: Ask clarifying questions, suggest common patterns, create a basic task and offer to refine.
Complex description: Suggest breaking into smaller tasks, focus on core functionality first, offer to create related tasks.
Missing technical details: Make reasonable assumptions, include multiple approaches, note areas needing user decisions.
Plan file not found: Check if path is a directory (look for plan.md within), suggest common PDD plan locations.
Invalid plan format: Identify missing sections, suggest running PDD to generate a proper plan, extract what's available.
All steps complete: Inform user, ask if they want a specific step anyway, suggest reviewing for new steps.
Other measured skills in the registry, with their headline benchmark lift.