Ralph Feature Setup Skill

Ralph Feature Setup

Interactive feature planning that creates ralph-ready tasks with dependencies.

The Job

Two modes:

Mode 1: New Feature

Chat through the feature - Ask clarifying questions
Break into small tasks - Each completable in one iteration
Create task_list tasks - Parent + subtasks with dependsOn
Set up ralph files - Save parent ID, reset progress.txt

Mode 2: Existing Tasks

Find existing parent task - Search or let user specify
Verify structure - Check subtasks have proper dependsOn
Set up ralph files - Save parent ID to parent-task-id.txt
Show status - Which tasks are ready, completed, blocked

Ask the user which mode they need:

Are you:
1. Starting a new feature (I'll help you plan and create tasks)
2. Using existing tasks (I'll set up Ralph to run them)

Step 1: Understand the Feature

Start by asking the user about their feature. Don't assume - ASK:

What feature are you building?

Then ask clarifying questions:

What's the user-facing goal?
What parts of the codebase will this touch? (database, UI, API, etc.)
Are there any existing patterns to follow?
What should it look like when done?

Keep asking until you have enough detail to break it into tasks.

Step 2: Break Into Tasks

Each task must be completable in ONE Ralph iteration (~one context window).

Ralph spawns a fresh Amp instance per iteration with no memory of previous work. If a task is too big, the LLM runs out of context before finishing.

Right-sized tasks:

Add a database column + migration
Create a single UI component
Implement one server action
Add a filter to an existing list
Write tests for one module

Too big (split these):

"Build the entire dashboard" → Split into: schema, queries, components, filters
"Add authentication" → Split into: schema, middleware, login UI, session handling
"Refactor the API" → Split into one task per endpoint

Rule of thumb: If you can't describe the change in 2-3 sentences, it's too big.

Step 3: Order by Dependencies

Tasks execute based on dependsOn. Earlier tasks must complete before dependent ones start.

Typical order:

Schema/database changes (migrations)
Server actions / backend logic
UI components that use the backend
Integration / E2E tests

Use dependsOn to express this:

Task 1: Schema (no dependencies)
Task 2: Server action (dependsOn: [task-1])
Task 3: UI component (dependsOn: [task-2])
Task 4: Tests (dependsOn: [task-3])

Parallel tasks that don't depend on each other can share the same dependency.

Step 4: Create Tasks

First, create the parent task:

task_list create
  title: "[Feature Name]"
  description: "[One-line description of the feature]"
  repoURL: "https://github.com/snarktank/untangle"

Save the returned task ID - you'll need it for subtasks.

Then, create subtasks with parentID and dependsOn:

task_list create
  title: "[Task title - action-oriented]"
  description: "[Detailed description with:
    - What to implement
    - Files to create/modify
    - Acceptance criteria
    - How to verify (typecheck, tests, browser)]"
  parentID: "<parent-task-id>"
  dependsOn: ["<previous-task-id>"]  // omit for first task
  repoURL: "https://github.com/snarktank/untangle"

Task description format:

Write descriptions that a future Ralph iteration can pick up without context:

Implement category name to ID mapping for expenses.

**What to do:**
- Create function mapExpenseCategoryNameToId(name, isChildExpense)
- Query item_category table with category_type filter
- Add alias mapping for common synonyms (rent → Rent or Mortgage)

**Files:**
- workflows/tools/upsert-expense.ts

**Acceptance criteria:**
- Function returns category ID for valid names
- Returns null for unknown categories
- npm run typecheck passes

**Notes:**
- Follow pattern from upsert-income.ts
- EXPENSE type for family, CHILD_EXPENSE for child

Step 5: Set Up Ralph Files

After creating all tasks, run the shared setup steps from "Final Setup (Required for Both Modes)" section.

This ensures:

Parent task ID is saved to scripts/ralph/parent-task-id.txt
Previous progress.txt is archived if it has content
Fresh progress.txt is created with Codebase Patterns preserved

Step 6: Confirm Setup

Show the user what was created:

✅ Ralph is ready!

**Parent task:** [title] (ID: [id])

**Subtasks:**
1. [Task 1 title] - no dependencies
2. [Task 2 title] - depends on #1
3. [Task 3 title] - depends on #2
...

**To run Ralph:**
```bash
./scripts/ralph/ralph.sh [max_iterations]
# or directly:
npx tsx scripts/ralph/ralph.ts [max_iterations]

To check status:

amp task list --parentID [parent-id] --limit 10

Mode 2: Setting Up Existing Tasks

If the user already has tasks created, help them set up Ralph to run them.

Find the parent task:

Ask the user:

What's the parent task? You can give me:
- The task ID directly
- A search term and I'll find it
- Or say "list recent" to see recent tasks

To search for tasks (always use limit to avoid context overflow):

task_list list
  repoURL: "https://github.com/snarktank/untangle"
  limit: 10

Verify subtasks exist:

Once you have the parent ID, check for subtasks (always use limit):

task_list list
  parentID: "<parent-task-id>"
  limit: 10

If no subtasks found, the parent task might BE the work (not a container). Ask:

This task has no subtasks. Is this:
1. A parent task with subtasks I should find differently?
2. The actual work task (I should create it as a parent with this as the first subtask)?

Check dependencies:

Review the subtasks and verify:

Do they have dependsOn set correctly?
Are there any circular dependencies?
Is the first task(s) dependency-free so Ralph can start?

If dependencies are missing, offer to fix:

These tasks don't have dependencies set. Should I:
1. Add dependencies based on their order?
2. Leave them parallel (Ralph picks any ready task)?

Set up ralph files:

Run the shared setup steps from "Final Setup (Required for Both Modes)" section below.

Show status:

✅ Ralph is ready to use existing tasks!

**Parent task:** [title] (ID: [id])

**Status:**
- ✅ Completed: 3 tasks
- 🔄 Ready to work: 2 tasks  
- ⏳ Blocked: 5 tasks (waiting on dependencies)

**Next task Ralph will pick:**
[Task title] - [brief description]

**To run Ralph:**
```bash
./scripts/ralph/ralph.sh [max_iterations]
# or directly:
npx tsx scripts/ralph/ralph.ts [max_iterations]

Final Setup (Required for Both Modes)

ALWAYS run these steps after creating tasks OR setting up existing tasks:

1. Save parent task ID:

echo "<parent-task-id>" > scripts/ralph/parent-task-id.txt

Verify it was saved:

cat scripts/ralph/parent-task-id.txt

2. Check if progress.txt needs archiving:

Read the current progress.txt:

cat scripts/ralph/progress.txt

Archive if:

It has content beyond the header (learnings from a previous feature)
The previous feature is different from the current one

Archive command:

DATE=$(date +%Y-%m-%d)
FEATURE="previous-feature-name-here"
mkdir -p scripts/ralph/archive/$DATE-$FEATURE
cp scripts/ralph/progress.txt scripts/ralph/archive/$DATE-$FEATURE/
echo "Archived to scripts/ralph/archive/$DATE-$FEATURE/"

3. Reset progress.txt:

Preserve useful Codebase Patterns from the previous run, then reset:

cat > scripts/ralph/progress.txt << 'EOF'
# Ralph Progress Log
Started: [current date]

## Codebase Patterns
[Copy any patterns from previous run that are still relevant]
---
EOF

4. Verify setup is complete:

# Confirm parent ID is saved
cat scripts/ralph/parent-task-id.txt

# Confirm progress.txt is reset
head -10 scripts/ralph/progress.txt

# List subtasks to confirm they exist
# (use task_list list with parentID)

Only after completing all 4 steps is Ralph ready to run.

Acceptance Criteria for Each Task

Each task description MUST include verifiable acceptance criteria:

Good criteria (verifiable):

"Add investorType column to investor table with default 'cold'"
"Filter dropdown has options: All, Cold, Friend"
"npm run typecheck passes"
"npm test -- path/to/test.ts passes"

Bad criteria (vague):

❌ "Works correctly"
❌ "Good UX"
❌ "Handles edge cases"

Always include:

npm run typecheck passes

For tasks with testable logic:

npm test passes

For UI/browser tasks:

For functional testing (checking behavior, not appearance):

Use Chrome DevTools MCP with take_snapshot to read page content (NOT screenshots)

take_snapshot returns the a11y tree as text that Ralph can read and verify
take_screenshot returns images which Ralph cannot interpret for functional checks
Only use screenshots when testing visual appearance/look and feel

For visual testing (checking appearance):

Use take_screenshot to capture and verify visual appearance

Example Conversation

User: set up ralph for adding expense tracking to the legal agent

Assistant: What feature are you building?

User: I want Grace (our legal chat agent) to be able to create, update, and delete expenses when users mention them in conversation.

Assistant: Got it - an expense upsert tool for the legal agent. A few questions:

Are there existing tools in the legal agent I should follow as a pattern?
Do expenses have different types (family vs child)?
Should it also list expenses or just CRUD?
Any task completion logic needed?

User: Follow upsert-income.ts. Yes family and child expenses (different tables). Yes list too. Mark APP_EXPENSES_ENTERED task complete.

Assistant: Perfect. Here's my proposed breakdown:

Create tool skeleton - Zod schema, basic structure
Category mapping - Map names like "rent" to category IDs
Amount parsing - Handle currency formats, allow $0
Duplicate detection - Check before creating
Create family expenses - Insert into expense table
Create child expenses - Insert into child_expense table
Update action - Modify existing expenses
Delete action - Remove expenses
List action - Show all expenses with totals
Task completion - Mark APP_EXPENSES_ENTERED done
Register tool - Add to legal chat workflow
UI progress indicator - Show progress during tool use
Unit tests - Test helper functions
E2E tests - Test via chat interface

Does this look right? Any tasks to add or remove?

User: Looks good, create the tasks

Assistant: [Creates parent task, then subtasks with dependencies, saves parent-task-id.txt]

✅ Ralph is ready!

Parent task: Legal Agent Expense Upsert Tool (ID: task-abc123)

Subtasks: 14 tasks created with dependencies

To run: ./scripts/ralph/ralph.sh 20

How Ralph Completes

When all subtasks are completed:

Ralph marks the parent task as completed
Ralph outputs <promise>COMPLETE</promise>
The ralph.sh/ralph.ts loop detects this and exits

Important: Ralph uses limit: 5 when querying tasks to avoid context overflow. If you have many subtasks, they'll be processed over multiple iterations.

Checklist Before Creating Tasks

[ ] Chatted through feature to understand scope
[ ] Each task completable in one iteration (small enough)
[ ] Tasks ordered by dependency (schema → backend → UI → tests)
[ ] Every task has "npm run typecheck passes" in description
[ ] UI tasks have browser verification in description
[ ] Descriptions have enough detail for Ralph to implement without context
[ ] Parent task ID saved to scripts/ralph/parent-task-id.txt
[ ] Previous run archived if progress.txt had content

Phase 2: The Execution Loop

Once setup is complete, Ralph runs the autonomous loop to implement tasks one by one.

Loop Workflow

0. Get the parent task ID

First, read the parent task ID that scopes this feature:

cat scripts/ralph/parent-task-id.txt

If this file doesn't exist, ask the user which parent task to work on.

Check if this is a new feature: Compare the parent task ID to the one in scripts/ralph/progress.txt header. If they differ (or progress.txt doesn't exist), this is a NEW feature - reset progress.txt:

# Build Progress Log
Started: [today's date]
Feature: [parent task title]

## Codebase Patterns
(Patterns discovered during this feature build)

---

This ensures each feature starts with a clean slate. Progress.txt is SHORT-TERM memory for the current feature only.

1. Check for ready tasks (with nested hierarchy support)

The task hierarchy may have multiple levels (parent → container → leaf tasks). Use this approach to find all descendant tasks:

Step 1: Get all tasks for the repo

task_list action: "list", repoURL: "<repo-url>", ready: true, status: "open", limit: 10

Important: Always use limit (5-10) to avoid context overflow with many tasks.

Step 2: Build the descendant set Starting from the parent task ID, collect all tasks that are descendants:

Find tasks where parentID equals the parent task ID (direct children)
For each child found, recursively find their children
Continue until no more descendants are found

Step 3: Filter to workable tasks From the descendant set, select tasks that are:

ready: true (all dependencies satisfied)
status: "open"
Leaf tasks (no children of their own) - these are the actual work items

CRITICAL: Skip container tasks that exist only to group other tasks. A container task has other tasks with its ID as their parentID.

2. If no ready tasks

Check if all descendant tasks are completed:

Query task_list list with repoURL: "<repo-url>" (no ready filter)
Build the full descendant set (same recursive approach as step 1)
If all leaf tasks in the descendant set are completed:
1. Archive progress.txt:
```
DATE=$(date +%Y-%m-%d)
FEATURE="feature-name-here"
mkdir -p scripts/ralph/archive/$DATE-$FEATURE
mv scripts/ralph/progress.txt scripts/ralph/archive/$DATE-$FEATURE/
```
2. Create fresh progress.txt with empty template
3. Clear parent-task-id.txt: echo "" > scripts/ralph/parent-task-id.txt
4. Commit: git add scripts/ralph && git commit -m "chore: archive progress for [feature-name]"
5. Mark the parent task as completed
6. Stop and report "✅ Build complete - all tasks finished!"
If some are blocked: Report which tasks are blocked and why

3. If ready tasks exist

Pick the next task:

Prefer tasks related to what was just completed (same module/area, dependent work)
If no prior context, pick the first ready task

Execute the task:

Use the handoff tool with this goal:

Implement and verify task [task-id]: [task-title].

[task-description]

FIRST: Read scripts/ralph/progress.txt - check the "Codebase Patterns" section for important context from previous iterations.

When complete:

1. Run quality checks: `npm run typecheck` and `npm test`
   - If either fails, FIX THE ISSUES and re-run until both pass
   - Do NOT proceed until quality checks pass

2. Update AGENTS.md files if you learned something important:
   - Check for AGENTS.md in directories where you edited files
   - Add learnings that future developers/agents should know (patterns, gotchas, dependencies)
   - This is LONG-TERM memory - things anyone editing this code should know
   - Do NOT add task-specific details or temporary notes

3. Update progress.txt (APPEND, never replace) - this is SHORT-TERM memory for the current feature:

[Date] - [Task Title]

Thread: [current thread URL] Task ID: [task-id]

What was implemented
Files changed
Learnings for future iterations:
- Patterns discovered
- Gotchas encountered
- Useful context


4. If you discovered a reusable pattern for THIS FEATURE, add it to the `## Codebase Patterns` section at the TOP of progress.txt

5. Commit all changes with message: `feat: [Task Title]`

6. Mark task as completed: `task_list action: "update", taskID: "[task-id]", status: "completed"`

7. Invoke the ralph skill to continue the loop

Progress File Format

# Build Progress Log
Started: [date]
Feature: [feature name]
Parent Task: [parent-task-id]

## Codebase Patterns
(Patterns discovered during this feature build)

---

## [Date] - [Task Title]
Thread: https://ampcode.com/threads/[thread-id]
Task ID: [id]
- What was implemented
- Files changed
- **Learnings for future iterations:**
  - Patterns discovered
  - Gotchas encountered
---

Note: When a new feature starts with a different parent task ID, reset progress.txt completely. Long-term learnings belong in AGENTS.md files, not progress.txt.

Task Discovery

While working, liberally create new tasks when you discover:

Failing tests or test gaps
Code that needs refactoring
Missing error handling
Documentation gaps
TODOs or FIXMEs in the code
Build/lint warnings
Performance issues

Use task_list action: "create" immediately. Set appropriate dependsOn relationships.

Browser Verification

For UI tasks, specify the right verification method:

Functional testing (checking behavior, not appearance):

Use Chrome DevTools MCP with take_snapshot to read page content

take_snapshot returns the a11y tree as text that can be read and verified
Use for: button exists, text appears, form works

Visual testing (checking appearance):

Use take_screenshot to capture and verify visual appearance

Use for: layout, colors, styling, animations

Quality Requirements

Before marking any task complete:

npm run typecheck must pass
npm test must pass
Changes must be committed
Progress must be logged

Stop Condition

When no ready tasks remain AND all tasks are completed:

Output: "✅ Build complete - all tasks finished!"
Summarize what was accomplished

Important Notes

Always use ready: true when listing tasks to only get tasks with satisfied dependencies
Always use limit: 5-10 when listing tasks to avoid context overflow
Each handoff runs in a fresh thread with clean context
Progress.txt is the memory between iterations - keep it updated
Prefer tasks in the same area as just-completed work for better context continuity
The handoff goal MUST include instructions to update progress.txt, commit, and re-invoke this skill

Agent Skills: Ralph Feature Setup

Skill Files