Create PRD from Deltas (Parallel Agent Orchestration)

Generate

prd.json

files from deltas that are ready to implement (status: ✓ Plan). This skill orchestrates parallel agent execution for efficient delta processing.

Note: This skill generates JSON files compatible with ralph-tui's JSON tracker plugin.

---

The Job

1. Load

   prd-task-creation

   skill for schema reference
2. Read

   docs/planning/DELTAS.md

   to find all deltas with status "✓ Plan"
3. Ask user which deltas to include (if multiple available)
4. Read

   docs/planning/DEPENDENCIES.md

   for cross-delta dependencies
5. Dispatch

   delta-task-splitter

   agents in parallel (one per delta)
6. Collect agent outputs from their responses (NOT from files they write)
7. Combine into unified prd.json with re-sequenced IDs and cross-delta dependencies
8. Output the assembled

   ./prd.json

Important: Do NOT start implementing. Just create the prd.json file.

---

Step 1: Load Schema Reference

Load the

prd-task-creation

skill using the Skill tool.

This provides the JSON schema reference for:

• Validating agent outputs
• Assembling the final prd.json structure
• Ensuring anti-patterns are avoided during combination

---

Step 2: Delta Discovery

Read the delta inventory and identify ready-to-implement deltas.

Filter Criteria

Look for deltas with

**Status**: ✓ Plan

in

docs/planning/DELTAS.md

.

Dependency Checking

Read

docs/planning/DEPENDENCIES.md

to understand any dependencies between deltas.

• If a delta depends on another that is NOT "✓ Plan", note this as a prerequisite
• Order deltas by dependency (prerequisites first)
• If no dependencies exist, order by delta ID

User Selection

If multiple deltas are ready, ask the user which to include:

Found N deltas ready to implement. Which would you like to include in the PRD?

   A. All ready deltas
   B. Only Easy deltas
   C. Only [category] deltas
   D. Specify delta IDs (e.g., "DLT-009, DLT-011")

---

Step 3: Dispatch Agents in Parallel

For each selected delta, dispatch a

delta-task-splitter

agent.

Preparing Agent Inputs

For each delta, prepare minimal input:

• delta_id: The delta identifier (e.g., "DLT-013")
• story_id_start: Estimated starting story number (will be re-sequenced after)

Note: Agents read their own spec, design, plan, and referenced documents using predictable file paths. This keeps orchestrator context minimal.

Parallel Dispatch

CRITICAL: Use the Task tool with multiple calls in a SINGLE message for parallel execution.

Example with 3 deltas:

Use Task tool 3 times in the SAME message:

Task 1:
  subagent_type: "delta-task-splitter"
  prompt: """
    delta_id: DLT-013
    story_id_start: 1
  """

Task 2:
  subagent_type: "delta-task-splitter"
  prompt: """
    delta_id: DLT-014
    story_id_start: 4
  """

Task 3:
  subagent_type: "delta-task-splitter"
  prompt: """
    delta_id: DLT-015
    story_id_start: 7
  """

The agents will run in parallel. Each agent reads its own documentation using predictable file paths.

---

Step 4: Collect and Combine Results

Collecting Agent Outputs

Each agent returns a JSON object in its response output:

{
  "delta_id": "DLT-XXX",
  "stories": [...],
  "story_count": N,
  "next_story_id": M
}

Important: Collect these outputs from the agent responses (Task tool results), NOT from files that agents may have written. The delta-task-splitter agents should return JSON data in their output, not write files.

Wait for all agents to complete and collect their outputs.

Re-sequencing Story IDs

Since agents run in parallel with estimated starting IDs, re-sequence after collection:

1. Order delta results by delta ID (or user-specified order)
2. Build an ID mapping: old ID -> new sequential ID
3. Assign sequential IDs: US-001, US-002, ... across all deltas
4. Update all

   dependsOn

   references using the ID mapping

Example:

Agent 1 (DLT-013) returned: US-001, US-002, US-003
Agent 2 (DLT-014) returned: US-004, US-005, US-006, US-007
Agent 3 (DLT-015) returned: US-008, US-009

After re-sequencing (if needed):
DLT-013: US-001, US-002, US-003
DLT-014: US-004, US-005, US-006, US-007
DLT-015: US-008, US-009

Resolving Cross-Delta Dependencies

From

docs/planning/DEPENDENCIES.md

, identify delta-level dependencies.

For each delta that depends on another:

1. Find the review story of the prerequisite delta (last story in that delta's sequence)
2. Add that story ID to the first story of the dependent delta's

   dependsOn

   array

Example: If DLT-015 depends on DLT-014:

• DLT-014's review story is US-007
• DLT-015's first implementation story (US-008) gets

  "dependsOn": ["US-007"]

Adjusting Priorities

After cross-delta dependencies are set:

1. Stories with no dependencies get priority 1
2. For stories with dependencies: priority = max(dependency priorities) + 1
3. This ensures topological ordering by priority

---

Step 5: Assemble Final JSON

Combine all stories into the final prd.json structure:

{
  "name": "[Derived from delta names or feature]",
  "branchName": "ralph/[kebab-case-feature-name]",
  "description": "[Summary of included deltas and their purpose]",
  "userStories": [
    // All stories from all agents, re-sequenced with cross-delta deps
  ]
}

Naming Conventions

• name: Describe what's being implemented (e.g., "Implement Phase 2 Deltas")
• branchName: Kebab-case, prefixed with

  ralph/

  (e.g., "ralph/phase-2-deltas")
• description: Brief summary of included deltas

---

Step 6: Output and Validation

Output Location

Assemble and write the final prd.json to

./prd.json

(or user-specified path).

Note: This is the ONLY file that should be written. The delta-task-splitter agents return data in their outputs, they do NOT write files.

Validation Checklist

Before saving, verify:

• JSON has flat structure (name, branchName, description, userStories at root)
• NO wrapper object (prd/tasks/metadata)
• Using "userStories" not "tasks"
• Using "passes" not "status"
• All story IDs are sequential (US-001, US-002, ...)
• All

  dependsOn

  references point to valid story IDs
• No circular dependencies
• Each delta has a review story at the end
• Quality gates appended to every story's acceptanceCriteria
• Cross-delta dependencies correctly resolved
• Priorities reflect dependency ordering

---

Running with ralph-tui

After creating prd.json:

ralph-tui run --prd ./prd.json

Ralph-tui will:

1. Load stories from prd.json
2. Select the highest-priority story with

   passes: false

   and no blocking dependencies
3. Generate a prompt with story details + acceptance criteria
4. Run the agent to implement the story
5. Mark

   passes: true

   on completion
6. Repeat until all stories pass

---

Error Handling

Agent Failure

If a

delta-task-splitter

agent fails:

1. Note which delta failed
2. Report the error to the user
3. Offer to retry that specific delta or continue without it

Invalid Agent Output

If an agent returns invalid JSON:

1. Attempt to parse and extract stories anyway
2. If unparseable, report error and offer to retry
3. Log the raw output for debugging

Dependency Cycle Detection

Before output, verify no circular dependencies exist:

1. Build a dependency graph from

   dependsOn

   arrays
2. Detect cycles using topological sort
3. If cycle detected, report affected stories and ask user how to resolve