/do — Structured Task Workflow
A Claude Code skill that turns any task into a systematic, quality-controlled workflow with external agent review.
Plan an implementation, get it reviewed by another AI agent, get your approval, implement with subagents, run automated QC, then present the results.
Quick Start
/do Add a dark mode toggle to the settings page
That's it. The skill auto-detects your project type and any available review agents. No configuration needed.
To resume an interrupted workflow from a saved plan:
/do --continue
How It Works
| Phase | What happens |
|---|---|
| 1. Plan | Claude researches the codebase and creates a step-by-step implementation plan with a complexity estimate |
| 2. Analyze | An external agent (Gemini, Codex, Claude, etc.) independently reviews the plan |
| 3. Approve | You review the plan + feedback and approve, reject, or request changes |
| 4. Implement | Auto-creates a do/<task-slug> git branch, then spawns subagents to implement the approved plan |
| 5. QC | Automated tests/builds run, then an external agent reviews the diff — safe suggestions are auto-applied, risky ones are surfaced for review |
| 6. Present | Summary of changes, branch name, QC results, and outstanding items for final approval |
Supported Agents
The skill can use any of these agents for plan and code review. All are optional — the skill detects what's available and falls back gracefully.
| Agent | Install | Type | Detection |
|---|---|---|---|
| Gemini CLI | npm i -g @google/gemini-cli | Cloud | which gemini |
| Codex CLI | brew install codex | Cloud | which codex |
| Ollama | brew install ollama | Local | which ollama |
| OpenRouter | Set OPENROUTER_API_KEY | Cloud | env var |
| Claude Code | npm i -g @anthropic-ai/claude-code | Cloud | which claude |
| Aider | pip install aider-chat | Cloud/Local | which aider |
| OpenAI-compatible | Set OPENAI_API_KEY | Cloud/Local | env var |
| GitHub Copilot | npm install -g @github/copilot-cli | Cloud | which copilot |
Agent availability is cached at ~/.claude/do-env.json. To force re-detection:
/do --refresh-env <task description>
Configuration Commands
Manage .claude/do-config.json without editing it manually:
/do config # show current config
/do config set provider copilot # use Copilot for both plan and code review
/do config set provider copilot,gemini # Copilot with Gemini as fallback
/do config set plan gemini # plan review only
/do config set code copilot:gpt-4o # code review only, specific model
/do config set model gpt-4.1 # update model on all API providers
/do config set qc test "npm test" # add/replace a QC command
/do config unset provider # revert to auto-detection
/do config reset # delete the config file
/do config refresh # clear agent detection cache
Both : and / work as the model separator — copilot:gpt-4o and copilot/gpt-4o are equivalent.
Flags
| Flag | Description |
|---|---|
--continue | Resume from a saved plan in .claude/plans/. Skips planning and jumps straight to approval. |
--refresh-env | Clear the agent detection cache and re-detect available agents before running. |
Installation
Plugin marketplace
/plugin marketplace add manaporkun/claude-plugins
/plugin install task
/reload-plugins
The skill is available as /task:do.
Release Management
Plugin versions are managed independently. The repository uses a generated Release Please config with:
- one root package for the overall marketplace release
- one Release Please package per
plugins/*directory - one post-processing step in the release workflow that syncs
.claude-plugin/marketplace.jsononto the release PR branch
Normal workflow for skill changes
If you change an existing skill under plugins/<name>/, you usually do not edit version numbers manually.
- Make your code changes.
- Commit with a conventional commit message scoped to the plugin, for example: