I built a CLI that scaffolds agentic workflows for Claude Code

#ai #claude #cli #webdev

Most developers using Claude Code work the same way: one long conversation, manual prompting between tasks, context that grows until the model loses track of what it was doing. It works — but it doesn't scale.

I built AgentKit to fix this. It's a CLI that generates the orchestration layer for any project: the files that tell Claude Code who to be, what to build, and how to hand off work between specialized agents — automatically.

How it works

You write a plain-language blueprint describing what you want to build:

# Todo App — Blueprint

## Goal
A simple web todo app with task creation, completion, and filtering.

## Tech constraints
- Next.js 14 + TypeScript + Tailwind
- localStorage for persistence, no backend

You run one command:

npx @patricksardinha/agentkit-cli init --blueprint PROJECT_BLUEPRINT.md

AgentKit generates four files:

CLAUDE.md — the standing brief every agent reads before starting
AGENT_WORKFLOW.md — a placeholder Claude Code fills during Phase 0
PLAYBOOK.md — the autonomous execution engine
README.md — project documentation extracted from your blueprint

Then you open Claude Code and type one instruction:

Read PLAYBOOK.md and execute the procedure.

What happens next

Claude Code runs in three phases automatically.

Phase 0 — Decomposition: it reads your blueprint and proposes a list of specialized agents. You validate before anything gets built.

Skills enrichment pause: Claude Code creates the agent folders and waits. You can drop API docs, schemas, or any context files into the relevant agent folder before execution starts.

Phase 1 — Execution: each agent runs in sequence, validates its success criterion, retries on failure, and escalates to you only when genuinely blocked.

A real example

I used AgentKit to build DevLog Desktop — a Tauri v2 app with local RAG via Ollama.

Result after one Claude Code session:

6 agents executed autonomously
68 tests passing
Windows installer produced (MSI + NSIS)
Full EN/FR i18n, dark mode, semantic search

Design choices

No AI inside the tool. AgentKit is purely structural — it generates markdown files. The intelligence lives in Claude Code, not in AgentKit. This means no API key, no costs, no vendor lock-in.

Phase 0 always runs. Most developers don't know how to optimally split work into agents. AgentKit delegates that decision to Claude Code every time, whether or not you provide a blueprint. You describe intent, Claude Code designs the structure.

Stack detection first, interactive fallback. AgentKit reads your project directory to detect the stack. If nothing is found (new empty project), it asks you interactively from a list of supported stacks: React, Next.js, Tauri, FastAPI, Express, Node.js.

Try it

# On any project (empty or existing)
npx @patricksardinha/agentkit-cli init

# With a blueprint
npx @patricksardinha/agentkit-cli init --blueprint PROJECT_BLUEPRINT.md

# Add a feature to an existing project
npx @patricksardinha/agentkit-cli add --feature "add dark mode"