Azure Lean Stack Skills
alexpizarro/azure-lean-stack-skills
Deploy Azure apps sem custo quando idle, com Bicep e CI/CD automático
Ver no GitHubPack de 14 skills compostas para Claude Code que permitem construir e deployar aplicações React + Azure Functions + SQL com custo zero quando ociosas. Inclui padrões de CI/CD branch-per-env, 37+ gotchas documentados e configurações de scale-to-zero prontas para produção.
Instalação
npx convez-ai alexpizarro/azure-lean-stack-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Scaffold de projetos React + Azure Functions + SQL com IaC Bicep
- Setup de CI/CD automatizado com branch-per-environment (local/test/prod)
- Otimização de custos de infraestrutura Azure com scale-to-zero defaults
- Resolução de gotchas conhecidos do Azure CLI e Azure Functions
Veredito

Skill pack maduro e opinionado que resolve problemas reais de deployment Azure que Claude Code enfrenta sozinho. Excelente para quem quer infraestrutura lean em produção sem overhead de configuração manual, mas requer familiaridade prévia com Azure e GitHub Actions.
README
A tradução para o português deste README ainda está na fila.
Azure Lean Stack
Azure apps that cost nothing when nobody's using them.
A Claude Code skill pack for building React + Azure Functions + SQL web apps the lean way — scale-to-zero defaults, branch-per-environment CI/CD, and 37+ documented gotchas so Claude Code can actually deploy to Azure on the first try.
Quick start
Prerequisites: Claude Code, Azure CLI (az login done), GitHub CLI (gh auth login done), Node.js 22.
# 1. Install the plugin
claude plugin install alexpizarro/azure-lean-stack-skills
# 2. Open an empty directory for your new project
mkdir my-azure-app && cd my-azure-app && claude
# 3. Ask Claude to scaffold:
# "Scaffold a new Azure web app. Org: acme. Project: taskapp. GitHub: acme/taskapp."
# 4. After files are generated, ask Claude to set up CI/CD:
# "Set up Azure OIDC and GitHub secrets for this repo."
# 5. Push to the test branch — first deploy in ~5 minutes
git push origin test
That's it. Idle cost: ~$0. Light-traffic cost: ~$5/month. Full cost table below.
Pin to a specific version for stable installs across the v2 → v3 jump:
claude plugin install alexpizarro/azure-lean-stack-skills@v2.1.0
Why this exists
I burned out on vibe coding. Lovable, v0, Bolt — they all let me ship something that looked like a real app in an afternoon. The moment I tried to take any of them past PoC, the wheels came off. Bloated bundles. Hardcoded values everywhere. State management nobody could maintain. Dead on arrival as a real-world app.
So I went back to what I knew worked: Microsoft Azure. Static Web Apps, Functions, SQL Serverless. Free tier on everything. The kind of stack you can hand to a client and they can keep running for a few dollars a month.
One problem: Claude Code couldn't deploy it.
Wrong Functions plan (Y1 instead of FC1). Deprecated app settings (FUNCTIONS_WORKER_RUNTIME is forbidden on FC1). SWA in the wrong region (australiaeast isn't supported — eastasia only). gpg: cannot open /dev/tty in CI. Forty-seven different ways the Azure CLI silently does the wrong thing.
So I built this skill pack. Azure Lean Stack. Every pattern in here is proven in a real production project — no fabricated examples. When Claude Code uses these skills, deployments work.
What you get
~$0 idle cost for the whole stack
~$5 monthly cost at light traffic
~14 composable Claude Code skills
~37 documented gotchas with verified fixes
~3 real production projects that prove every pattern
Local → test → prod, one branch per tier — main runs a fully-offline local stack (Docker SQL + Azurite, no Azure cost); push to test → test env; push to production → prod env. No GitHub Environments to configure, no deploy approvals to set up, no manual terraform apply from a laptop.
Free tier by default — Static Web Apps Free, SQL Serverless with auto-pause, Container Apps scale-to-zero, Application Insights with daily quota cap, Storage with lifecycle rules. Pay for what you actually use.
Scale up when you need to — every default is tunable. When traffic justifies it, flip the parameter file and redeploy.
The stack
| Layer | Default | Cost when idle |
|---|---|---|
| Frontend | React 19 + Vite 6 on Static Web Apps Free | $0 |
| API | Azure Functions v4 (Node 22) — SWA managed functions | $0 |
| Database | Azure SQL Serverless GP_S_Gen5_1, auto-pause at 15 min | ~$0.10/mo (1GB cap) |
| Background jobs | Container Apps Jobs with minReplicas: 0 | $0 |
| Scheduled tasks | Logic Apps Consumption | $0.22/mo at 5-min cadence |
| Azure Communication Services Email | $0 (first 100/day free) | |
| Observability | Application Insights with dailyQuotaGb: 1 | <$2/mo |
| Storage | Blob Standard_LRS + lifecycle rules (Hot→Cool@60d→Cold@180d) | ~$0.02/GB/mo |
| CI/CD | GitHub Actions with OIDC (no client secrets) | $0 |
A typical low-traffic prototype runs at under $5/month total.
What Claude does under the hood
When you ask Claude to scaffold a new app, it routes through these sub-skills in order:
scaffolding-azure-bicep-infrastructure— generatesinfra/,.github/workflows/, parameter filesconfiguring-azure-oidc-for-github-actions— creates service principals + federated credentials + GitHub secretsmanaging-azure-sql-migrations— sets up the migration systemdeploying-azure-static-web-apps— generates the React + Functions codeapplying-azure-cost-guardrails— audits the Bicep for cost regressions before first deploy
You don't have to invoke these by name. Describe what you want — "scaffold an Azure app with SQL and storage", "set up OIDC", "add a recurring scheduler" — and Claude picks the right sub-skill via orchestrating-azure-deployments.
Deploy workflow
Always the same three commands:
# Deploy to test
git checkout test && git merge main -m "Promote to test: <summary>" && git push origin test
# Deploy to production
git checkout production && git merge test -m "Release to prod: <summary>" && git push origin production
# Back to local dev
git checkout main
main is never connected to an Azure environment. It exists for local dev only.
Need a new isolated environment? git checkout -b acme-demo, set up one OIDC federated credential bound to the branch, push. You have a fully isolated Azure environment in 10 minutes.
The 16 skills
| Skill | When Claude uses it |
|---|---|
orchestrating-azure-deployments | The router — picks the right sub-skill for any Azure task |
scaffolding-azure-bicep-infrastructure | New project, generating infra/ and workflows |
configuring-azure-oidc-for-github-actions | Service principals, federated credentials, GitHub secrets |
managing-azure-sql-migrations | Idempotent SQL migrations run via sqlcmd in CI |
deploying-azure-static-web-apps | SWA + managed functions code conventions |
deploying-fc1-flex-consumption-functions | Timer/queue/AI workloads on standalone Function Apps |
deploying-azure-container-apps | Long-running servers, Jobs, sidecars, multi-app envs |
scheduling-with-azure-logic-apps-consumption | Recurring HTTP triggers and lightweight Power-Automate-style flows |
developing-azure-apps-locally | Fully-offline local stack (Docker SQL + Azurite) — the main-branch "try" tier |
optimizing-azure-blob-storage-cost | Lifecycle rules, CORS for SAS+Range, tier ageing |
adding-azure-communication-services-email | Transactional email (100/day free) |
instrumenting-azure-app-insights | Workspace-based App Insights with daily cap + alerts |
scaffolding-multi-tenant-azure-apps | One RG per tenant on a shared subscription |
applying-azure-cost-guardrails | Consumption-first defaults + a Bicep auditor script |
diagnosing-azure-deployment-failures | 37+ gotcha catalogue with verified fixes |
curating-azure-deployment-learnings | Captures field learnings and promotes recurring ones to gotchas |
See RECIPES.md for working patterns lifted from real projects.
Branch-per-environment — the deployment model
This is the single most important pattern in the pack. One branch in your repo = one tier. The flow is local (main) → test → production.
main → the local "try" tier — fully-offline Docker stack (SQL + Azurite), no Azure
test → deploys to {org}-{project}-rg-test
production → deploys to {org}-{project}-rg-prod
acme-demo → deploys to {org}-{project}-rg-acme-demo
You iterate on main against a real local SQL + blob stack (see developing-azure-apps-locally), then promote to test, then production.
Why this works:
- Branch state IS environment state. Always. The branch is the source of truth for what's running.
- OIDC federated credential subject is
repo:owner/repo:ref:refs/heads/{branch}— Azure literally won't authenticate a workflow run from the wrong branch. Test SPs cannot deploy to prod, even if a workflow tries. - No GitHub Environments to configure. No deploy approvals. No protected-branches dance.
- New environment = new branch. 10 minutes from "we need a demo for next Tuesday" to "here's the URL."
configuring-azure-oidc-for-github-actions automates the SP + federated-credential setup for any new branch.
How this complements Microsoft's Azure Skills plugin
Microsoft published their official Azure Skills plugin with 25 skills + 200+ MCP tools. They cover the live Azure surface — querying resources, running cost analysis, doing RBAC audits.
This pack complements that with opinionated, proven patterns:
| Microsoft's Azure Skills | Azure Lean Stack (this) | |
|---|---|---|
| Scope | All of Azure, generically | The lean-cost SWA + Functions + SQL + ACA stack |
| Strength | Live data via Azure MCP | Battle-tested CI/CD + gotcha catalogue |
| Best for | "What's running in my subscription?" | "Why did my deploy fail?" / "How do I keep this under $5/month?" |
| Installation | microsoft/azure-skills | alexpizarro/azure-lean-stack-skills |
Install both. Claude Code uses Microsoft's for live diagnostics; Azure Lean Stack for the deployment patterns that aren't in Microsoft's official surface.
Architecture decisions (non-negotiable)
- Branch-per-environment.
main= dev only;testandproduction(and any other branch) map 1:1 to Azure resource groups via branch-scoped OIDC. - SWA managed functions by default. Free tier. HTTP only. Switch to FC1 only when you need timer/queue/AI triggers.
- SQL Serverless.
GP_S_Gen5_1, auto-pauses at 15 min idle. - OIDC auth. Never a client secret in CI.
- JSON parameter files.
.parameters.jsonnot.bicepparam(needs inline--parametersoverrides). - SWA location is
eastasia. Hard-coded;australiaeastdoesn't supportMicrosoft.Web/staticSites. - Every pattern is proven in a real project. Nothing in this pack is fabricated — if no shipping project uses it, it doesn't get added.
Production readiness — known gaps
Azure Lean Stack is consumption-priced and optimised for fast iteration. For real production workloads, address these:
| Priority | Gap | Upgrade path |
|---|---|---|
| 1 | No authentication | Entra ID on SWA, or APIM with JWT validation |
| 2 | SQL password auth | Managed Identity with db_datareader/db_datawriter |
| 3 | Over-scoped SPs | Scope from subscription to resource group |
| 4 | Public SQL access | Private endpoint + VNet integration |
| 5 | No observability by default | Set deployObservability: true (skill: instrumenting-azure-app-insights) |
| 6 | SQL cold starts on idle | Upgrade from serverless to provisioned tier |
| 7 | No WAF | Azure Front Door + WAF policy |
Contributing
The skill pack grows from real field experience. The workflow:
- Hit a problem deploying a real Azure project
- Capture it via
curating-azure-deployment-learnings:bash skills/curating-azure-deployment-learnings/scripts/capture-learning.sh - When the same problem appears in 2+ projects (or is high-severity), promote it:
bash .../promote-to-gotchas.sh <tag> - PR with the updated gotcha + the Bicep/script change
No fabricated patterns. Every skill in this pack must be proven in at least one shipping project. If you want to add a new skill, link to the project that already uses it.
License
MIT — use it, fork it, ship clients on it.