What I learn about cloud, architecture and agents, written to be reusable. All of it free to read.
From empty repo to a live product in a day: the plan, the shortcuts, and what breaks when you go too fast.
→Mega-prompts, no CLAUDE.md, no tests, no commits: the traps that turn the agent into a liability.
→A clean end-to-end pipeline: stages, cache, artifacts, environments and automatic deployment.
→Zero infra, per-PR previews, global edge: what Vercel saves you — and the day you must leave it.
→Amplify, EC2, ECS, or serverless: the four paths to host a SaaS on AWS, and how to choose.
→Leaked secrets, unprotected API routes, IDOR, XSS: the holes I find on almost every app.
→Your own GitLab on a server: Docker, backups, CI runners, and the sizing to plan for.
→LCP, CLS, INP: the concrete levers — images, fonts, server components — for an app that flies.
→Sessions vs JWT, httpOnly cookies, OAuth, and where to put the checks: an auth foundation that holds.
→Stop abuse without hurting real users: token bucket, Redis, and what to limit first.
→No more accidentally committed .env: where secrets live, how to inject them, how to rotate them.
→Know it broke before the customer does: structured logs, metrics, alerts that don’t cry wolf.
→An untested backup isn't a backup. Retention, real restore drills, and the 3-2-1 rule.
→How I structure a growing project: when a monorepo helps, when it becomes a burden.
→Coding in the terminal with an agent: the mental model, the commands that matter, and how to keep it on the rails.
→The real playbook: tagging, rightsizing, Savings Plans vs Spot — without breaking production.
→NAT Gateway, egress, forgotten snapshots, oversized EBS, logs: the hidden costs that bleed scale-ups.
→A decision tree based on load, budget and team size — not hype.
→From "works on my machine" to reproducible and observable: IaC, CI/CD, health checks, rollback.
→Load balancer, auto-scaling, read replicas, cache — sized right to absorb a spike.
→Ground on real data, function calling, guardrails: escape the believable fiction.
→When the agent should ask vs decide alone, and how to interrupt then resume a durable run.
→The architecture that survives a traffic spike: EC2 for the site, RDS for the database, S3 + Lambda for WebP-optimized product photos.
→Get the agent out of the black box: instrument every step, measure cost, and score it with an LLM-as-a-judge.
→