Tasks to PRD Converter (with Smart Routing)

Convert user stories from tasks.md into prd.json format WITH automatic routing classification.

Note: This command is typically auto-run by the /relentless.tasks skill. Manual execution is only needed if you edit tasks.md by hand after generation.

The Job

⚠️ CRITICAL: This skill requires executing the relentless convert CLI command. Do NOT generate prd.json manually!

1. Locate tasks.md in the current feature directory
2. Validate story structure, dependencies, and TDD compliance
3. Preview conversion (show story count, dependency chain)
4. EXECUTE the CLI command: relentless convert relentless/features/NNN-feature/tasks.md --feature <feature-name>
5. Validate the generated prd.json HAS routing metadata
6. Display routing summary (cost estimates per story)
7. Report next step

Why Routing Matters

Routing is the core value of Relentless:

• Simple tasks use cheap models (haiku, flash) - saves money
• Complex tasks use SOTA models (opus, gpt-5) - ensures quality
• Estimated costs BEFORE execution - no surprises
• Compare estimated vs actual after execution - learn and improve

Without routing, you lose the intelligent cost optimization that makes Relentless valuable.

Step 1: Find Feature Directory

Look in relentless/features/ for the most recent feature:

ls -la relentless/features/

Verify:

• tasks.md exists in the feature directory
• spec.md exists (for context)
• Feature directory follows naming convention (NNN-feature-name)

If tasks.md doesn't exist, suggest running /relentless.tasks first.

Step 2: Validate tasks.md Structure

The CLI now automatically validates tasks.md before conversion. You can also validate manually:

relentless validate relentless/features/NNN-feature/tasks.md

Validation Checks

The validator checks for:

• Each story has unique ID (US-XXX format)
• Each story has acceptance criteria (after filtering)
• Dependencies reference existing stories
• No circular dependencies
• No underscore format in dependencies (US_001 should be US-001)

Automatic Filtering (with Warnings)

The following criteria are automatically filtered during conversion:

• Standalone file paths: `src/file.ts` → Add context: "src/file.ts contains X"
• Section markers: **Files:** → Move outside acceptance criteria section
• Short text: Less than 3 characters

Example Valid Story

### US-001: Create User Registration

**Description:** As a new user, I want to register so I can access the app.

**Acceptance Criteria:**
- [ ] POST /api/register endpoint exists
- [ ] Email validation works
- [ ] Password is hashed
- [ ] `src/auth/register.ts` contains the registration handler
- [ ] Unit tests pass
- [ ] Integration test passes

**Dependencies:** None
**Phase:** Foundation
**Priority:** 1

Example Validation Output

Validating tasks.md...

⚠️  2 validation warning(s):
  [FILTERED_CRITERIA]
    2 acceptance criteria will be filtered during conversion

  📝 2 criteria will be filtered during conversion
     US-003: "`src/queue/types.ts`"
     US-004: "**"

Step 3: Run Conversion (with Routing)

⚠️ IMPORTANT: Execute this bash command using the Bash tool. Do NOT generate prd.json manually!

relentless convert relentless/features/NNN-feature/tasks.md --feature <feature-name>

Replace NNN-feature with the actual feature directory name (e.g., 001-user-auth).

Options

Expected Output

Converting tasks.md...

Classifying 6 stories for routing (mode: good)...
  US-001: simple -> claude/haiku-4.5 ($0.0012)
  US-002: medium -> claude/sonnet-4.5 ($0.0034)
  US-003: complex -> claude/opus-4.5 ($0.0156)
  US-004: simple -> claude/haiku-4.5 ($0.0010)
  US-005: medium -> claude/sonnet-4.5 ($0.0028)
  US-006: expert -> claude/opus-4.5 ($0.0234)

Total estimated cost: $0.0474

Created relentless/features/NNN-feature/
  prd.json - 6 stories
  Routing metadata included
  prd.md - from tasks.md
  progress.txt - progress log

Step 4: Validate Routing Output

After conversion, verify the prd.json has routing metadata:

cat relentless/features/<feature>/prd.json | jq '.userStories[0].routing'

Expected Routing Structure

{
  "complexity": "simple",
  "harness": "claude",
  "model": "haiku-4.5",
  "mode": "good",
  "estimatedCost": 0.0012,
  "classificationReasoning": "Task classified as simple (confidence: 85%)..."
}

CRITICAL: If routing is missing, the conversion was run with --skip-routing. Re-run without that flag.

Step 5: Report Summary

After successful conversion, provide this summary:

## Conversion Complete

**Feature:** NNN-feature-name
**Stories:** 6 total

### Routing Summary

| Story  | Complexity | Harness/Model        | Est. Cost |
|--------|------------|----------------------|-----------|
| US-001 | simple     | claude/haiku-4.5     | $0.0012   |
| US-002 | medium     | claude/sonnet-4.5    | $0.0034   |
| US-003 | complex    | claude/opus-4.5      | $0.0156   |
| US-004 | simple     | claude/haiku-4.5     | $0.0010   |
| US-005 | medium     | claude/sonnet-4.5    | $0.0028   |
| US-006 | expert     | claude/opus-4.5      | $0.0234   |

**Total Estimated Cost:** $0.0474

### Files Created

- `relentless/features/<feature>/prd.json` - PRD with routing
- `relentless/features/<feature>/prd.md` - Copy of tasks.md

### Next Steps

1. Review routing decisions (adjust mode if needed)
2. Run `/relentless.analyze` to check consistency
3. Run `relentless run --feature <name> --mode good`

Complexity Classification

The classifier determines complexity based on:

Mode-Model Matrix

Troubleshooting

"Tasks.md not found"

Run /relentless.tasks first to generate tasks.md from spec.md and plan.md.

"No routing metadata"

The conversion was run with --skip-routing. Re-run:

relentless convert relentless/features/<feature>/tasks.md --feature <feature-name>

"Validation failed"

The tasks.md has format issues. Run validation to see details:

relentless validate relentless/features/<feature>/tasks.md

Common issues:

• DUPLICATE_STORY_ID: Two stories have the same ID
• MISSING_DEPENDENCY: A story depends on non-existent story ID
• CIRCULAR_DEPENDENCY: Stories form a dependency loop

"Criteria will be filtered"

Some acceptance criteria don't meet format requirements. Fix by:

• Adding context to file paths: `src/file.ts` → "src/file.ts contains X"
• Moving section markers outside acceptance criteria
• Writing longer, more descriptive criteria

"Invalid dependency"

A story references a non-existent story ID. Check the Dependencies: line.

"Circular dependency detected"

Stories reference each other in a cycle. Example: US-001 depends on US-002, and US-002 depends on US-001.

"Dependency uses underscore format"

Use dashes in story IDs: US-001 not US_001.

Notes

• Always run conversion WITH routing (default behavior)
• Routing metadata enables intelligent cost optimization
• Review complexity classifications before running
• Use --mode free for testing/experimentation
• Use --mode good (default) for production work
• Use --mode genius for critical/complex features