https://github.com/badlogic/claude-commands
Global Claude Code commands and workflows
https://github.com/badlogic/claude-commands
Last synced: about 2 months ago
JSON representation
Global Claude Code commands and workflows
- Host: GitHub
- URL: https://github.com/badlogic/claude-commands
- Owner: badlogic
- Created: 2025-07-06T20:03:55.000Z (4 months ago)
- Default Branch: main
- Last Pushed: 2025-08-11T00:30:29.000Z (3 months ago)
- Last Synced: 2025-08-11T02:31:33.174Z (3 months ago)
- Size: 46.9 KB
- Stars: 422
- Watchers: 14
- Forks: 25
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- awesome-claude-code - **claude-commands**
README
# Todo Implementation System
Human-agent collaborative workflows for transforming vague todos into implemented features in ~150 LOC of Markdown.
This workflow is meant for engineers who want to stay in control of and informed about their code base, using agents to do the tedious work, and their brain for the hard parts.
**Two workflows available:**
- **`todo-worktree.md`**: Full task isolation using git worktrees, separate branches, and incremental commits. Best for complex features or when working on multiple tasks in parallel.
- **`todo-branch.md`**: Simpler workflow on current branch with single commit at completion. Best for quick fixes, single-task focus, or when worktrees aren't desired.
**Problems this solves**:
- Breaking vague ideas into well defined problem descriptions and concrete implementation steps
- Preventing agents from making unauthorized changes or expanding scope
- Efficient context usage by providing agents only task-relevant information
- Full isolation of feature work (with worktrees) or simple staging workflow (without)
- Resuming interrupted work without losing context or progress
- Traceable feature development history with meaningful commits and pull requests
**Problems this does not solve**:
- "Claude, build me Uber for dogs"
- Slop in your code base generated by an army of unsupervised agents in the cloud
- Merge conflicts
**Why this approach**:
- Zero dependencies, pure markdown + git + shell commands
- Easy to understand, modify, and transfer between any coding agent
- No vendor lock-in, maximum transparency and simplicity
## How it works
Both workflows follow the same state machine pattern with slight variations. Read [`todo-worktree.md`](./todo-worktree.md) or [`todo-branch.md`](./todo-branch.md) to understand the full flow and adapt it to your needs.
At its core, each workflow is a state machine that guides the agent and the user through distinct phases:
- **`INIT`**: Sets up the project description (`todos/project-description.md`) and checks for orphaned tasks from previous sessions offering them for resumption
- **`SELECT`**: The human chooses a task from `todos/todos.md`
- **`REFINE`**: The agent researches the codebase and collaborates with the human to create a detailed implementation plan
- **`IMPLEMENT`**: The agent executes the plan step-by-step, with the human reviewing, revising and approving each change
- **`COMMIT`**: The work is finalized - either as a PR (todo-worktree.md) or a single commit (todo-branch.md)
**Key concepts:**
- **Todo list**: `todos/todos.md` - Simple markdown file with your vague ideas, one per line or bullet. Just jot down ideas/issues quickly here.
- **Task transformation**: Selected todos become detailed `task.md` files with description, implementation plan, and status tracking
- **Project context**: `project-description.md` provides reusable, extensible context about the codebase
- **Research preservation**: Agent analysis is saved to `analysis.md` for reference
- **Isolation strategies**:
- `todo-worktree.md`: Git worktrees with separate branches
- `todo-branch.md`: Current branch with progressive staging
See also [prompt-as-code](https://mariozechner.at/posts/2025-06-02-prompts-are-code/) workflow.
## Installation
### Claude Code
1. Choose your workflow:
- `todo-worktree.md` for worktree-based isolation (recommended for most users)
- `todo-branch.md` for simpler single-branch workflow
2. Put your chosen file in `~/.claude/commands` (global) or `project/.claude/commands` (project)
- If you put it in your project, ensure you commit it before starting work on any todos
3. In your project: `mkdir -p todos && echo "- add dark mode" > todos/todos.md`
### Other agents
1. Choose your workflow (see descriptions above)
2. Put your chosen file in a location accessible to your agent
3. In your project, create your first todo: `mkdir -p todos && echo "- add dark mode" > todos/todos.md`
## Usage
The core idea is to load your chosen workflow as a system prompt or initial instruction for your agent. The agent will then run the workflow defined in the prompt and guide you through it. Like in a text adventure.
### Claude Code
In your project root directory:
```bash
# For worktree workflow
claude --dangerously-skip-permissions "/todo"
# For no-worktrees workflow
claude --dangerously-skip-permissions "/todo-branch"
# The --dangerously-skip-permissions flag allows the agent to execute
# shell commands (git, cd, mkdir) as defined in the workflow without
# user approval. Omitting it will require you to manually approve
# every shell command.
```
Claude will guide you through the workflow.
### Other agents
1. Run your agent in your project root directory
2. Give the contents of your chosen workflow file to the agent as the first user message
The agent will guide you through the workflow.
## Caveats
**LLMS are unreliable**: LLMs occasionally ignore their programming when tasks get complex. Keep your tasks small and focused for best results.
**`todo-worktree.md`**: The worktree isolation lets you run agents in parallel, even across machines, but orphaned task detection only works if all agents run on the same machine.
**`todo-branch.md`**: Works on your current branch with uncommitted changes throughout the process. Make sure your working tree is clean before starting. Consider switching to a feature branch to avoid working directly on main.
**It's just a prompt** Much of this could be proper code instead of a prompt. But then it wouldn't be as portable or easy to tweak. Your agent can help convert the workflows to a real program if you prefer, calling into an agent for steps that require (interactive) inference.
**Concurrency**: Multiple humans editing `todos/todos.md` simultaneously creates perfect race conditions. In team projects, coordinate who's working on what so you don't end up with duplicate PRs for the same feature. Replacing `todos/todos.md` with one GitHub issue per vague todo is not a solution by the way. Think it through.
**Army of agents**: Distributed task management across machines would need a central server, which defeats the simplicity. Left as an exercise for distributed systems enthusiasts and their silicon assistants.