Skip to content

About

Multi-agent setup built around a single orchestrator, that delegates work to specialized agents for prompt refinement, requirements, design, planning, implementation, review, and logging.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

Feature Workflow

Move fast without collapsing everything into one oversized chat context.

Feature Dev Workflow is a multi-agent setup built around a single orchestrator, feature-dev, that delegates work to specialized agents for prompt refinement, requirements, design, planning, implementation, review, and logging.

The repository ships two variants:

  • .github/agents is the active VS Code-oriented setup.
  • portable/agents is a runtime-neutral baseline ready to be dropped into provider-specific folders. See the Claude Code setup section below for concrete instructions.

Why This Setup Works

A single-agent workflow collapses all concerns into one growing context. As scope expands, quality degrades. This workflow keeps each agent focused on a narrow responsibility, so context stays tight and outputs stay reliable.

The orchestrator, feature-dev, owns routing and approvals. Specialized subagents handle the rest. Backend work skips the design phase entirely. Only phases that actually run are logged.

The VS Code variant is operationally stronger because it uses native tools like vscode/askQuestions for approval gates. This is not just a UX detail: it prevents unnecessary premium-request consumption during checkpoints and keeps the workflow strict without burning tokens on avoidable back-and-forth.

Agent Architecture

Orchestrator

  • feature-dev.agent.md
    • Owns the full workflow.
    • Handles routing, approvals, and phase transitions.
    • Treats FE/BE classification as an internal routing decision, not a standalone logged phase.

User-Facing Agents

  • prompt-builder.agent.md

    • Refines raw requests into approved prompts.
    • Explicitly classifies requests as frontend, backend, or full-stack.
  • documenter.agent.md

    • Saves phase checkpoints, final feature summaries, notes, and ADR-style documentation.

Subagents

Skill

Workflow

At runtime, the workflow executes in this order:

  1. Prompt intake
  2. Internal routing as frontend, backend, or full-stack
  3. Requirements creation
  4. Optional design phase (frontend and full-stack only)
  5. Planning
  6. Implementation
  7. Code review
  8. Logging and closeout

Key behaviors:

  • prompt-builder is optional. If the request is already well-scoped, feature-dev accepts it directly.
  • FE/BE/full-stack detection always happens, but it is not logged as a dedicated phase.
  • Backend-only work skips designer.
  • Every approved phase that actually runs is checkpointed by documenter, followed by a final synthesis pass at closeout.

Flow Diagram

flowchart TD
    U[User Request] --> FD[feature-dev]
    FD --> PB[prompt-builder optional]
    FD --> RT{Internal routing\nFE/BE/full-stack}
    PB --> RT
    RT --> RC[requirements-creator]
    RC --> DG{Needs design?}
    DG -->|Yes| DS[designer]
    DG -->|No| PL[planner]
    DS --> PL
    PL --> IM[implementer]
    IM --> CR[code-reviewer]
    CR --> DC[documenter]
Loading

Quick Start

VS Code Setup

Prerequisites:

  • VS Code with GitHub Copilot custom agents support
  • GitHub Copilot access enabled for chat and agents
  • A git repository for the target project
  • Agent discovery from .github/agents

If your environment does not automatically discover repository-local agents, verify the relevant VS Code settings such as chat.agentFilesLocations.

git clone https://github.com/Elverle/feature-workflow
cd feature-workflow

Open the repository in VS Code. The agents are versioned directly in .github/agents, so the project is already structured for repository-based sharing.

Claude Code Setup

portable/agents contains runtime-neutral agent files with no VS Code-specific tooling. These are designed to be dropped into Claude Code's custom agent directory without modification.

Step 1 — Clone the repository

git clone https://github.com/Elverle/feature-workflow
cd feature-workflow

Step 2 — Copy the portable agents into your target project

Claude Code discovers agents from the .claude/agents/ folder at the root of your project. Copy the portable agents there:

cp portable/agents/*.md /path/to/your-project/.claude/agents/

If .claude/agents/ does not exist yet, create it first:

mkdir -p /path/to/your-project/.claude/agents/

Step 3 — Copy the frontend-design skill (optional)

If your project includes frontend work and you want the designer subagent to use the skill, copy it as well:

mkdir -p /path/to/your-project/.claude/skills/frontend-design/
cp .github/skills/frontend-design/SKILL.md /path/to/your-project/.claude/skills/frontend-design/

Step 4 — Start the workflow

Open Claude Code in your project and invoke the orchestrator:

@feature-dev Build a dashboard for tracking payout reconciliation and add the backend endpoints that power it.

Approval gates in Claude Code

The VS Code variant uses vscode/askQuestions for gated approvals. In Claude Code, the portable agents fall back to standard chat-turn confirmations: the orchestrator will ask for explicit approval at each checkpoint before continuing. This is slightly less efficient in token usage but functionally equivalent.

Differences from the VS Code variant

Capability VS Code Claude Code
Approval gates Native vscode/askQuestions Chat-turn confirmations
Agent discovery .github/agents/ .claude/agents/
File write tools Native VS Code filesystem Claude Code bash/edit tools
Token efficiency at gates Higher Slightly lower

Use the Right Entry Point

  • Use feature-dev for the full orchestrated workflow.
  • Use prompt-builder when the request is still too vague to scope.
  • Use documenter when you want to document decisions or notes outside the main workflow.

Examples

Backend-Only Request

Skip prompt building. I need an internal batch job that reconciles supplier payouts.

What happens:

  1. feature-dev accepts the user request as the approved prompt.
  2. It classifies the feature as backend.
  3. It invokes requirements-creator.
  4. It skips designer.
  5. It invokes planner, implementer, code-reviewer, and documenter.

Full-Stack Request

Build a dashboard for tracking payout reconciliation and add the backend endpoints that power it.

What changes:

  1. The workflow classifies the feature as full-stack.
  2. designer is invoked after approved requirements.
  3. The generated design artifacts are passed into planner and then into implementer.

Generated Artifacts

The workflow writes feature artifacts under feature/feature-{number}/.

Typical outputs include:

  • prompt.md when prompt-builder ran
  • requirements.md
  • implementation-plan.md
  • 00-feature-summary.md
  • 01-prompt-builder-summary.md when the prompt phase ran
  • 02-requirements-summary.md
  • 03-designer-summary.md when the design phase ran
  • 04-planner-summary.md
  • 05-implementer-summary.md
  • 06-code-review-summary.md
  • prototypes/design-preview.html when the design phase ran
  • decisions/ and notes/ for additional feature documentation

At repository level, the workflow can also maintain feature-index.md.

Repository Structure

.
|-- README.md
|-- .github/
|   |-- agents/
|   |   |-- feature-dev.agent.md
|   |   |-- prompt-builder.agent.md
|   |   |-- documenter.agent.md
|   |   |-- requirements-creator.subagent.agent.md
|   |   |-- designer.subagent.agent.md
|   |   |-- planner.subagent.agent.md
|   |   |-- implementer.subagent.agent.md
|   |   `-- code-reviewer.subagent.agent.md
|   `-- skills/
|       `-- frontend-design/
|           `-- SKILL.md
|-- docs/
|-- portable/
|   |-- agents/
|   `-- docs/
`-- feature/
    `-- feature-{number}/

Notes:

  • feature/ is a generated output area used by the workflow at runtime.
  • docs/ is available for project documentation and examples.
  • portable/ contains the runtime-neutral agent copies for non-VS Code environments.

About

Multi-agent setup built around a single orchestrator, that delegates work to specialized agents for prompt refinement, requirements, design, planning, implementation, review, and logging.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors