Turboflow：为 Claude Code 打造的单管线插件，用 10 个技能、5 个智能体、1 个 hook 取代混乱配置

You have 6 agents in ~/.claude/agents/, 4 custom commands, 3 rule files that contradict each other, and a planner-actually-good.md you wrote at 2 AM. Claude picks whichever it finds first. You can't remember which is current. Neither can Claude.

Turboflow replaces all of it with a single opinionated pipeline: 10 skills, 5 agents, 1 hook. One system to install, nothing to maintain.

Install

Requires Claude Code 2.1.288 or later (claude --version).

claude plugin marketplace add nicodiansk/turboflow
claude plugin install turboflow@turboflow

Restart Claude Code, then start building:

/turboflow:brainstorm I want to build a CLI tool that manages git worktrees

Every skill announces itself, does one job well, and hands you the next step:

Other install methods

Local development (per-session):

claude --plugin-dir ./turboflow

Update:

claude plugin update turboflow@turboflow

Upgrading from turbocharge 2.x

3.0.0 renamed the plugin and its marketplace from turbocharge to turboflow, so every /turbocharge:<skill> is now /turboflow:<skill>. Claude Code follows the plugin rename on its own, but an existing install keeps its old marketplace key, turbocharge. Pick one path:

Quick — keep the old marketplace key:

claude plugin marketplace update turbocharge
claude plugin install turboflow@turbocharge

Clean (recommended) — match the install id above:

claude plugin marketplace remove turbocharge
claude plugin marketplace add nicodiansk/turboflow
claude plugin install turboflow@turboflow

Then run /reload-plugins in any open session. Note that marketplace remove also uninstalls the plugin and deletes its saved options and data.

Either way, update personal rules, permissions, and scripts that mention turbocharge:* or /turbocharge:<skill>. Full notes are in the CHANGELOG.

The Pipeline

brainstorm → story → plan → build → review → ship
                                ↑
                              debug

Enter at any step. Each skill gates the next. debug loops under build when bugs surface.

wrap stands alone — run it any time to capture session state, decisions, and a resume prompt for your next session.

atlas stands alone — generates ATLAS.md domain map from the actual codebase. Pre-loaded every session for zero-tool-call navigation.

Why a Pipeline

• One source of truth. No competing planner-v2.md files, no rules that contradict each other. Install once, update once.
• Each step gates the next. Requirements before stories, stories before plans, plans before code. The skill you skipped is usually the bug you ship.
• Right model for each role. Builders, reviewers, and researchers run on Sonnet — fast and focused. Your orchestration session runs whatever model you've configured, so you can reach for a more capable one on the planning and judgment calls.
• Drop in anywhere. Already have a plan? Start at build. Hit a bug? debug. Wrapping up? wrap. You're never forced through the whole chain.

How Build Works

By default each task runs builder (Sonnet) → self-review → commit — one spawn per task. For security-sensitive or unfamiliar code, the opt-in --reviewed flag adds a single Sonnet task-reviewer that loads the diff once and returns Spec + Quality verdicts; on issues it loops back to the builder for up to two cycles (2–6 spawns per task).

Every N tasks (default 3, set with --checkpoint=N; --checkpoint=0 runs straight through), the pipeline checkpoints with you for feedback. After all tasks, chain to /turboflow:review for the final holistic assessment.

Multi-track mode — independent tasks can run in parallel with coordinated builders. Requires Agent Teams (experimental):

claude config set env.CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS 1

Dispatched by skills. You never invoke them directly.

All agents have memory: project for persistent codebase knowledge across sessions.

Hooks

What the hook does

hooks/session-start.sh runs whenever a session starts, resumes, or is cleared or compacted. It is read-only: it prints text to stdout and does nothing else. It writes no files, changes no settings or permissions, and sends nothing over the network itself. What it prints becomes part of the session context, so it reaches the model like any file Claude reads.

Quick Start Examples

# Full pipeline from scratch
/turboflow:brainstorm I want to build a REST API for managing bookmarks

# Jump to planning with existing requirements
/turboflow:plan docs/plans/my-feature-stories.md

# Debug a specific issue
/turboflow:debug The auth middleware is rejecting valid tokens

# Wrap up your session
/turboflow:wrap

See examples/ for sample outputs from each stage.

What to Remove After Installing

Turboflow replaces scattered config. /turboflow:setup handles this, but in short:

• Agents in ~/.claude/agents/ (planner, code-reviewer, tdd-guide, session-wrappers)
• Commands in .claude/commands/ for story-authoring, task-breakdown, session-wrap
• Rule files that reference competing agents

Project Structure

turboflow/
├── .claude-plugin/
│   ├── plugin.json              # Plugin manifest
│   └── marketplace.json         # Distribution config
├── skills/                      # 10 skill definitions
│   └── <skill-name>/SKILL.md
├── agents/                      # 5 agent definitions
│   └── <agent-name>.md
├── hooks/
│   ├── hooks.json               # Hook registration
│   └── *.md                     # Hook content files
├── scripts/
│   ├── validate.sh              # Plugin health check
│   └── tests/                   # Content-shape tests
├── examples/                    # Sample pipeline outputs
├── settings.json
├── CHANGELOG.md
└── README.md

Validate

./scripts/validate.sh

License

MIT — see LICENSE.

_{Built by @nicodiansk — one pipeline to replace them all.}