flisk-recipe-starters
Start from a ready-made Flisk starter app instead of an empty project when the request matches one - restaurant (reservations, kitchen, menu, floor), SaaS (dashboard, billing, checkout), online store,
Files1
View on GitHubSKILL.md.claude/skills/flisk-recipe-starters/SKILL.md
GitHubRecipe: start from a starter app
A starter is a working multi-service app in the flisk-sh GitHub organisation (swarm, security, super-agent, langgraph and tool-loop ship inside the CLI and need no clone or Docker). Starting from one is faster than building from an empty folder, and the user can change it afterwards. Use one when the request clearly matches a row below; use only one starter per project.
| The user asks for | Template | Stacks | What it ships |
|---|---|---|---|
| restaurant, reservations, kitchen, menu, bistro, diner | restaurant |
guest, kitchen, operations |
reservation app and API, floor app, kitchen API and worker, menu API (two websites) |
| SaaS, dashboard, billing, subscription, checkout | saas |
app, billing |
dashboard app and API, checkout app, billing API, a migrator per API |
| shop, store, storefront, ecommerce, catalogue | ecommerce |
store |
Vue storefront, catalog API and its migrator |
| AI agent stack, agent with tools and a database | agent-stack |
agents |
agent, its migrator, MCP tools server |
| multi-agent swarm, coordinator and workers, task queue with retries | swarm |
none | coordinator and three workers as local processes (no Docker); retries a task whose worker died |
| security agents, code scanner, vulnerability triage | security |
none | scanner, triage and reporter agents as local processes (no Docker) |
| super agent, planner with tools and sub-agents, deep research | super-agent |
none | planner, MCP tool server, researcher, writer and checker as local processes (no Docker); plans with Claude when ANTHROPIC_API_KEY is set |
| LangGraph, agent graph, durable agent, resumable runs | langgraph |
none | a LangGraph StateGraph agent (plan, research, write, check) with checkpoints that resume after a crash, an MCP tool server, and a setup job that installs its pinned packages; local processes (no Docker) |
| plain agent loop, tool calling, Anthropic SDK, no framework | tool-loop |
none | the Anthropic SDK tool runner over an MCP tool server, with a local mock model so it runs offline; local processes (no Docker) |
| user management, admin portal, accounts, audit | user-management |
admin, compliance, portal |
admin app and API, portal app and API, audit API and worker |
| the smallest example, orders | example |
shop |
orders app, API and worker |
erp also exists: it is a 107-service scale test, not an app. Do not use it for a user's request.
Steps
flisk doctor
flisk project <template> --path <folder> --yes
cd <folder>
flisk up --dry-run # check the services and URLs
flisk up <stack> --json & # or omit <stack> to start every stack
flisk networks --raw
The starters ship .flisk/project.json, so flisk up and flisk stacks work right after cloning; flisk up generates the Tiltfile and refreshes the runtime assets itself. Do not run flisk project --yes in the clone unless flisk doctor asks for it (Generated Runtime Assets).
Then change the starter to what the user asked for: rename things in the app code, add resources with flisk resource ... --stack <stack>, and follow the flisk-build skill to check every URL and report them.
Guardrails
- Cloning needs network access to github.com. If it fails, fall back to the matching recipe (flisk-recipe-shop, flisk-recipe-saas, flisk-recipe-ai-agent) on an empty project.
- A starter with several stacks can be large on a small machine. Start only the stack the user needs first (
flisk up <stack>), and ask before starting all of them. - Tell the user which starter you used and that it came from github.com/flisk-sh.