Multi-stage workflow orchestration for AI coding assistants. Automate complex development tasks — feature development, bug fixes, code reviews, and refactoring — through structured, sequential agent pipelines.

An installable skill that turns your AI coding assistant into a workflow engine. Instead of giving your assistant one big prompt and hoping for the best, workflow-run breaks complex tasks into sequential stages — each executed by a focused agent, with context automatically passed between them.

Think of it as a task pipeline for AI agents: define stages like analyze, plan, implement, and review, and the orchestrator runs them in order with built-in failure handling and context management.

Works with any AI coding assistant that supports the skills format (Claude Code, Copilot CLI, Gemini CLI, and others via skills.sh).

npx skills add https://github.com/ananyasaxena/skills --skill workflow-run

Multi-stage workflow orchestrator that executes sequential stages with automatic context passing between them. Useful for complex tasks that benefit from structured decomposition into discrete phases.

• Workflow type resolution — automatically selects the best workflow type for your goal, or dynamically composes one if none fit
• Context passing — each stage receives a summary of prior stages, with sliding-window compression to stay within budget
• Failure handling — intelligent retry strategies that distinguish read-only stages (analysis, review) from side-effecting stages (implementation)
• Sub-workflows — stages can trigger child workflows, up to 3 levels of nesting
• Dynamic composition — when no built-in type fits, composes a custom workflow with guardrails (2-7 stages, must start with analysis, must end with review)
• Interactive workflow builder — create new workflow types through a guided Q&A, no JSON editing required

• Automate code review — analyze a branch diff, run a thorough review, and produce an actionable summary in one command
• AI-driven feature development — go from requirements to implementation to review without manual handoffs between stages
• Automated bug diagnosis and fix — reproduce, diagnose root cause, implement a fix with regression test, and verify
• Safe refactoring with AI — incremental refactoring with test verification at each step and behavioral change detection
• Custom development workflows — define your own multi-stage pipelines for deployment, migration, documentation, or any repeatable process

Run a workflow:

/workflow-run <goal>
/workflow-run feature-development Add user authentication with OAuth
/workflow-run bug-fix Login fails when email contains a plus sign
/workflow-run code-review Review changes on this branch before merge
/workflow-run refactoring Extract shared validation logic into a module

Create a new workflow type:

/workflow-run create
/workflow-run create my-workflow
/workflow-run create deploy-pipeline "Multi-environment deployment with canary stages"

The create command walks you through building a workflow interactively — it asks what the workflow is for, what to call it, what stages it needs, and what each stage should do. Everything is optional upfront; it derives what it can from your answers and only asks for what's missing. It then generates the JSON, shows you a draft for approval, and saves it to workflow-types/.

Workflow types are JSON files in workflow-types/. You can create them with the create command above, or write them by hand:

{
  "version": 1,
  "name": "my-workflow",
  "description": "What this workflow does and when to use it",
  "stages": [
    {
      "id": "analyze",
      "name": "Analyze",
      "prompt": "Analyze the goal: {GOAL}",
      "agent_type": null
    },
    {
      "id": "execute",
      "name": "Execute",
      "prompt": "Execute based on the analysis.",
      "agent_type": null
    }
  ]
}

Stage fields:

• id — unique identifier within the workflow
• name — human-readable stage name
• prompt — instructions for the stage executor (use {GOAL} placeholder for the user's goal)
• agent_type — optional, specifies a specialized agent type (e.g., code-reviewer); defaults to general-purpose

Any AI coding assistant that supports the skills plugin format:

• Claude Code — npx skills add
• Copilot CLI — via skills discovery
• Gemini CLI — via skill activation
• Any assistant using skills.sh

MIT