Installs into .claude/skills of the current project.
Are you the author of Tf Start?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/marblo-app-tf-start)
---
name: tf-start
description: Create tasks from the PRD and spawn agents to kick off the project. Use after /tf-plan.
disable-model-invocation: true
allowed-tools: Bash, Read, Glob, Grep, Write, Edit
argument-hint: [프로젝트명]
---
# Marblo Project Kickoff
> Run this after `/tf-plan` has finalized the PRD + task plan.
> If there is no PRD, point the user to `/tf-plan` first.
---
## ⛔ Required rule: Marblo MCP only
> **Never use the Claude Code built-in tools (TaskCreate, TaskList, TaskUpdate, TaskGet).**
> Use **only the Marblo MCP tools**:
>
> - Bulk-create tasks: `create_tasks_bulk` (create all tasks at once in a single call)
> - Create a single task: `create_task`
> - Lookup tasks: `get_all_tasks`, `get_available_tasks`
> - Work tasks: `claim_task`, `update_task_status`, `add_activity`, `submit_for_review`
> - Other: `check_feedback`, `get_task_activities`, `get_agent_skill`, `get_task_dependencies`
>
> **Do not register tasks individually with Claude Code's built-in TaskCreate.** They won't show on the Marblo dashboard.
> **Using `create_tasks_bulk` registers all tasks at once in a single call.**
---
## Pre-flight check
Confirm before starting:
1. **PRD check**: confirm a finalized plan exists in `docs/PRD.md` or the prior conversation
- If none: guide "Please plan first with `/tf-plan`" and stop
- If present: read the PRD and extract the task list
2. **Environment check**:
- Confirm Marblo MCP connection: test with a `get_all_tasks` call
- Confirm the project directory exists
- If existing tasks are present, check for conflicts
3. **Final user confirmation**:
```
📋 Tasks to create: {N}
📦 Project: {project_name}
TASK-001: [title] (backend, priority: 5)
TASK-002: [title] (backend, priority: 4, depends_on: TASK-001)
...
Create and start these as-is?
```
---
## Phase 1: Confirm project name + bulk-create tasks
### 1-1. Confirm project name
Use the project name stated in the PRD. If none, confirm with the user.
**This project name goes consistently into every task's `project` field.**
### 1-2. Convert task list → `create_tasks_bulk` JSON
Convert the PRD's task plan into **a single `create_tasks_bulk` call**.
Required fields per task:
- `title`: a clear title
- `description`: concrete work + completion criteria
- `role`: backend / frontend / test / devops
- `priority`: 5(urgent) ~ 1(low)
- `depends_on`: dependencies (TASK-NNN format)
- `scope`: file paths to modify (conflict prevention)
- `project`: **the project name locked in Phase 0** (same for all tasks)
### 1-3. Run the bulk creation
**Always call Marblo MCP's `create_tasks_bulk` once to create all tasks at once.**
```
⛔ Wrong way: call Claude Code's TaskCreate 13 times individually
✅ Right way: call Marblo MCP's create_tasks_bulk once
```
### 1-4. Verify the creation result
- Whether all succeeded
- Whether the dependency mapping is correct
- Confirm dashboard registration with `get_all_tasks`
---
## Phase 2: Spawn agents + start work
1. Load each role's skill file with `get_agent_skill`:
- backend → `skills/backend_agent.md`
- frontend → `skills/frontend_agent.md`
- test → `skills/test_agent.md`
- devops → `skills/devops_agent.md`
2. Use `get_available_tasks` to see tasks that can be handled immediately.
- tasks with no dependencies, or whose dependencies are already satisfied
3. As Team Leader, spawn the agents:
- pass each agent its skill file + task info
- separate file areas by scope → parallel work possible
4. Work loop:
```
for each available task:
1. claim_task (with agent_id)
2. update_task_status → IN_PROGRESS
3. do the actual coding work
4. add_activity to log progress (file creation, test results, etc.)
5. submit_for_review
6. check the next available task
```
5. Report during work:
```
🚀 Project kickoff complete
━━━━━━━━━━━━━━━━━━━
Tasks created: {N}
Ready to start: {M}
Agents spawned: backend, frontend
Currently in progress:
- TASK-001: DB schema design (Backend Agent)
Waiting:
- TASK-002: User API (waiting for TASK-001)
- TASK-003: Main UI (waiting for TASK-002)
```
---
## Phase 3: Monitoring
Keep checking while work is in progress:
1. On task completion → the unblocked next task starts automatically
2. When a REVIEW comes up → notify the user
3. When a FAILED occurs → analyze the cause + report to the user
4. Check PM feedback with `check_feedback` along the way
> If the user wants to stop midway, point them to `/tf-hold`.