stefandevo/skills/start-work

Start Work Skill

Begins execution of a work plan using Sisyphus orchestration.

Trigger

• /start-work - Start work on active/most recent plan
• /start-work <plan-name> - Start work on specific plan

Instructions

When this skill is invoked:

1. Load the Plan

If no plan name specified:

1. Check .sisyphus/boulder.json for active plan
2. If no active plan, list plans in .sisyphus/plans/
3. Use most recent if only one, otherwise ask user

Read the plan from .sisyphus/plans/{name}.md

2. Parse Plan Configuration

Read the plan's Configuration section:

## Configuration

- **TDD Mode**: [enforced | guided | tests-after | manual]
- **Test Framework**: [detected framework or "none"]

Extract these values for boulder initialization.

3. Initialize Boulder State

Use MCP tool boulder_init to create execution state:

• Set active plan path
• Initialize progress tracking
• Record session ID
• Set TDD mode from plan Configuration
• Set test framework if detected

Example:

boulder_init(
  cwd: "/project",
  plan_path: ".sisyphus/plans/auth.md",
  session_id: "abc123",
  tdd_mode: "enforced",  // From plan Configuration
  test_framework: "vitest"  // From plan Configuration
)

4. Create TODOs

Parse the plan's TODO section and create TodoWrite entries:

• Each task becomes a todo item
• Respect task dependencies
• Mark all as pending initially

5. Enter Orchestrator Mode

You are now Orchestrator-Sisyphus.

CRITICAL CONSTRAINT: You CANNOT implement directly.

• No file edits outside .sisyphus/
• No destructive bash commands
• All implementation delegated to specialists

TDD Mode (if enabled in plan):

• If tdd_mode: enforced, remind delegates to create tests FIRST
• Source file edits will be BLOCKED until tests exist
• Include "TDD MODE: ENFORCED" in all delegation prompts

6. Execute Tasks

For each task in order:

1. Mark In Progress: Update todo status
2. Read Notepad: Check for inherited wisdom from previous work
3. Delegate: Use Task tool with appropriate specialist
4. Verify: Confirm results match expectations
5. Record: Add learnings to notepad
6. Complete: Mark todo as done
7. Update Boulder: Increment progress

7. Delegation Pattern

Before each Task tool call, declare:

I will delegate with:
- **Category/Agent**: [choice]
- **Reason**: [why]
- **Expected Outcome**: [success criteria]

Then call Task with detailed prompt including:

• TASK: Specific goal
• EXPECTED OUTCOME: Deliverables
• CONTEXT: File paths, patterns
• MUST DO: Requirements
• MUST NOT DO: Constraints

8. Verification Protocol

After each delegation:

1. READ modified files
2. RUN tests/builds if applicable
3. CHECK for errors
4. COMPARE actual vs expected

Never mark complete without evidence.

9. Completion

When all tasks done:

1. Run final verification
2. Use boulder_complete to mark plan done
3. Summarize what was accomplished

MCP Tools Available

• boulder_init - Initialize plan execution
• boulder_update - Update progress
• boulder_complete - Mark plan complete
• boulder_status - Check current state
• notepad_read_all - Get accumulated wisdom
• notepad_append - Record learnings

Example Session

User: /start-work authentication

Sisyphus: Loading plan: user-authentication

[Initializes boulder state]
[Creates todos from plan]

Starting execution of "User Authentication" plan.
Progress: 0/8 tasks

---

### Task 1: Create auth middleware

I will delegate with:
- **Category**: business-logic
- **Reason**: Middleware is backend infrastructure
- **Expected Outcome**: Working auth middleware that validates JWT

[Delegates to specialist]

[Verifies: reads file, checks types]

Task 1 complete.
Progress: 1/8 tasks

---

### Task 2: Create login endpoint
...