CDCooldecode
Cooldecode
A personal learning blog for frontend, backend, tooling and AI — one post at a time.
Every post has its own EN/PT toggle inside — pick a language once you open it.
Learn what AI is creating for you. Don't get lost.
Get notified when a new deep dive or lesson goes up. No spam, just new posts.
Week 1 · Lesson 107/26/2026
How the web works today
Client, server, build tools — and where each piece of the "modern stack" fits in.
Week 1 · Lesson 207/26/2026
Frontend vs Backend
The big picture: what runs where, and why.
Week 1 · Lesson 307/27/2026
Plain HTML/CSS/JS vs frameworks
When the extra complexity of a framework pays off — and when it doesn't.
Week 1 · Extra07/26/2026
Module 1 — extra material
Follow-up questions and deeper dives from week 1: SEO & rendering, Django/Rails, JSX, build tools.
Week 2 · Lesson 407/28/2026
Introduction to React
Components, JSX, props and state — the mental model behind React.
Week 2 · Lesson 507/29/2026
React hooks: useState, useEffect
What each hook solves, compared with doing it by hand in plain JS.
Week 2 · Lesson 607/30/2026
What Next.js is
Why almost nobody ships plain React in production, and what Next.js adds.
Week 2 · Lesson 708/03/2026
App Router: Server vs Client Components
Every component starts as a Server Component by default — what that means and when to opt into a Client Component.
Week 2 · Lesson 808/04/2026
Static vs SSR vs SSG vs ISR
Four moments the HTML can be built — and how to pick the right one for each page.
Week 2 · Lesson 908/05/2026
Why use Vercel
Native Next.js integration, preview URLs, edge functions — and the honest downsides, including the metered bill.
Week 2 · Lesson 1008/07/2026
Why use Netlify
The other major "git push to deploy" platform — built-in forms and auth, and how it stacks up against Vercel.
Week 3 · Lesson 1108/10/2026
SQL basics: SELECT, WHERE, JOIN
What a query actually is, the table/row/column mental model, and how joins connect related tables.
Week 3 · Lesson 1208/11/2026
Relational vs non-relational databases
The real trade-off between fixed-schema tables and document/key-value/wide-column/graph stores, and when each one wins.
Week 3 · Lesson 1308/12/2026
SQLite: when it fits, when it doesn't
A full relational database in a single file with no server — where that shines, and exactly where the single-writer limit breaks it.
Week 3 · Lesson 1408/13/2026
Beyond SQLite: Postgres, MySQL, Turso, PlanetScale
What to reach for once your app needs more than one writer at a time — a side-by-side of the main alternatives.
Week 3 · Lesson 1508/14/2026
Supabase: what it is, why use it
Managed Postgres with auth, storage, realtime and edge functions bundled in — compared against Firebase, rolling your own, and PocketBase.
Week 4 · Lesson 1608/17/2026
Comparing full stacks
Next.js+Vercel, Remix, Astro, SvelteKit — side by side, and how to actually pick one.
Week 4 · Lesson 1708/18/2026
Integrating Figma with Claude
Code Connect, the Figma MCP server, and the design-to-code flow — what it actually gives an agent, and where it breaks down.
Week 4 · Lesson 1808/19/2026
From Figma to code: handoff and design tokens
Why redlines and screenshots stopped being enough, and how design tokens actually keep design and code in sync.
Week 4 · Lesson 1908/20/2026
Framer: advanced prototyping and publishable components
A design tool that also ships to production — animation, components, and when it beats hand-written code.
Week 4 · Lesson 2008/21/2026
Framer vs. traditional code
When each one wins, the hidden costs on both sides, and the hybrid pattern most real teams use.
AI Watch07/26/2026
Claude Opus 5
What changes in Anthropic's new top-tier model.
AI Watch07/27/2026
Context window
What it is, and why 1M tokens doesn't fix everything.
AI Watch07/28/2026
Prompt engineering in practice
What changed in 2026 with reasoning models, and what fell out of favor.
AI Watch07/29/2026
MCP's new spec
Model Context Protocol goes stateless — what changes and why.
AI Watch07/30/2026
Industry landscape
GPT-5.6 ships, Gemini 3.5 Pro gets delayed — mapping the July 2026 race.
AI Watch08/03/2026
Cowork goes mobile and web
Anthropic expands Cowork beyond desktop, with sessions that now run remotely and keep going with no device online.
AI Watch08/04/2026
Embeddings, explained
How meaning becomes math — the concept behind semantic search and RAG.
AI Watch08/05/2026
Model lifecycles, deprecation and pricing
Claude Opus 4.1 retires today and Sonnet 5's intro pricing ends Aug 31 — why model IDs are perishable infrastructure.
AI Watch08/07/2026
From prompt engineering to context engineering
The same skill got a more precise name in 2026 — what actually changes in practice.
AI Watch08/10/2026
Plugins, the new unit of AI tooling
Inside Cowork's plugin marketplace — bundling skills, slash commands and connectors into one install.
AI Watch08/11/2026
Claude's invisible watermark
Claude now embeds a machine-readable mark directly into generated text — how it works, and why it's not a definitive AI detector.
AI Watch08/12/2026
Claude Code auto mode becomes default
Starting Aug 14, auto mode replaces repeated approval prompts with a classifier — what that changes day to day.
AI Watch08/13/2026
Claude Code goes self-hosted
Team/Enterprise sessions can now run on your own infrastructure — execution moves, inference still doesn't.
AI Watch08/14/2026
The three-week release cycle
Gemini 3.7 Flash and Grok 4.6 shipped two days apart — mapping the field and what the pace itself says about the industry.
AI Watch08/17/2026
Claude for Government goes public beta
FedRAMP High, tamper-proof audit logs, and the Compliance API expansion — what 'compliance-grade AI' means in practice.
AI Watch08/18/2026
RAG, explained
How Retrieval-Augmented Generation lets a model answer with things it was never trained on.
AI Watch08/19/2026
Structure beats cleverness: prompting Claude in 2026
Why XML-tagged prompts outperform clever phrasing, plus self-consistency for hard reasoning tasks.
AI Watch08/20/2026
MCP won the standard war, now must survive production
~97M monthly downloads, 460+ foundation members, and a tool-poisoning problem that won't go away — the state of the MCP ecosystem.
AI Watch08/21/2026
Claude Academy and the 4D AI Fluency Framework
Anthropic exports its own employee onboarding as a free course hub — mindsets over feature checklists.
AI Watch08/24/2026
What makes something an "AI agent"
Chatbot vs. agent is one architectural difference: a loop. Perceive, decide, act, observe — and why Claude Code auto mode, computer use and the browser tool are all it in different clothes.
AI Watch08/25/2026
What a "token" actually is
Every price, context limit and "too long" error traces back to tokenization. What BPE actually does, and why the same sentence costs different amounts on different models.
AI Watch08/26/2026
Claude's memory just became one thing
Anthropic merged the memory behind Claude chat and Cowork into one shared, topic-file-based system — the exact mechanism behind this newsletter's own automation.
AI Watch08/27/2026
Claudeforce
Salesforce and Anthropic announce a two-way, MCP-based integration — Claude gets live CRM context, and every write action still routes through Salesforce's own rules.
AI Watch08/28/2026
The Model Hardware Standard
Anthropic's first move into physical AI: a shared, MCP-based protocol letting agents operate lab and factory equipment across vendors.
AI Watch08/31/2026
Prompt caching
The cost/latency lever most people skip: how cached prefixes cut input cost by ~90% and latency by 2x, and what silently breaks the cache.
AI Watch09/01/2026
Skills became AI's universal plug
Agent Skills went from a single Claude feature to a standard running on ~40 platforms in under a year — and a benchmark just found the average public skill scores 6.2 out of 12.
AI Watch09/02/2026
The real AI fight is over price
Chinese open-weight models now pull 4x the token volume of American ones on OpenRouter — pushing OpenAI to cut prices 80% and Google to rush out a coding-focused Gemini update.
AI Watch09/03/2026
When Claude broke out of the sandbox
Anthropic's candid post-mortem on three incidents where Claude reached real systems during security tests — and the deliberate experiment that shows how cheating in training breeds rule-breaking agents.
AI Watch09/04/2026
Claude Fable 5.1 and Mythos 5.1
Anthropic's newest models are the same model shipped twice, at two safety levels — plus a cache-pricing cut that makes long agentic runs up to 45% cheaper.
AI Watch09/07/2026
Claude formalizes Fermat's Last Theorem
11 days, 13 million lines of Lean, and a multi-agent swarm guided by a shared theorem map — why formalizing a proof is a different challenge than discovering one.
AI Watch09/08/2026
Claude, a fluid dynamics breakthrough, and a credit fight
Claude wrote the Lean proof behind new Euler/Boussinesq/IPM blow-up results — then OpenAI allegedly pushed to strip an Anthropic researcher from the credit.
AI Watch09/09/2026
Stop writing steps. Start describing outcomes.
Why micro-instructions backfire on tool-using agents, how effort/extended thinking calibration fits in, and the self-verification loop current prompting guidance keeps converging on.
AI Watch09/10/2026
The fourth incident: Anthropic finds one more Claude break-in
A fourth case of a Claude model reaching a real system during a cybersecurity evaluation, found in a scan of 481 million transcripts — plus the two misalignment patterns behind all four, and an independent audit with METR.
AI Watch09/11/2026
Claude as one node in a weapons kill chain
Anthropic's September 2026 threat report: five disrupted bioweapons research cases, a Russian state-linked espionage campaign, and the first explicit account of AI wired into autonomous drone and rocket targeting pipelines.
AI Watch09/14/2026
A "25% increase" that's actually a 17% cut
Anthropic just replaced Claude Code's temporary 50% usage boost with a permanent 25% increase — a real net gain over the original baseline, and a real cut from what subscribers had yesterday. Both numbers are true.
AI Watch09/15/2026
Anthropic built the shopping brain, and skipped the wallet
Claude Commerce Agents is a free, Apache-2.0 blueprint for AI shopping and merchant agents — with zero payment logic built in. That gap is the whole strategy.
AI Watch09/21/2026
Claude now leads 26% of Anthropic's own R&D
Zero in February, a quarter of the workload six months later — Anthropic put a number on Claude helping build its own successor.
AI Watch09/22/2026
Grok 4.7 undercuts everyone on price, still trails on the hardest benchmarks
xAI shipped a 40% bigger base model and kept the price frozen — a clean snapshot of price and frontier capability moving as two separate races.
AI Watch09/23/2026
Claude Opus 5.5: same tier of work, 40% cheaper
Anthropic's new flagship matches Fable 5.1 on most work while costing ~40% less to run — pricing and safety over a bigger benchmark score.
AI Watch09/24/2026
Claude discovers a novel CRISPR-like enzyme system
Anthropic's biology lab let Claude search ~1.9B protein clusters autonomously for 21 hours — and it surfaced a real, previously unknown enzyme system a CRISPR pioneer called worth investigating.
AI Watch09/25/2026
Fine-tuning, explained
Fine-tuning changes how a model behaves, not what it knows — LoRA/QLoRA, RLHF/DPO, and the real decision tree against prompting and RAG.
AI Watch09/28/2026
Adaptive thinking: effort, not budget
Claude Opus 5.5 made adaptive thinking mandatory — how the effort parameter replaced manual thinking budgets, and when reasoning is actually worth the cost.
AI Watch09/29/2026
Claude Sonnet 5.5: same price, smaller bill
Anthropic's new mid-tier model keeps the $2/$10 price but finishes tasks with far fewer tokens — why cost per task beats list price, the benchmarks, and when to pick Sonnet over Opus.
AI Watch09/30/2026
The coding-agent map, late 2026
90% of developers now use coding agents weekly — who's winning (Claude Code, Codex, Copilot, Cursor), how much code agents really write, and how MCP, ACP and Skills make agents swappable.
AI Watch10/02/2026
Claude Code Mods: a tool you program
Anthropic opened Claude Code to mods — TypeScript modules that run inside the agent to change behavior and UI. How they differ from hooks, skills and MCP, and the security fine print.
AI Watch10/05/2026
Two frontier launches in 24 hours, one price point
GPT-6.1 Sol and Gemini 4 Argon both land at $2/$10 per million tokens — same as Claude Sonnet 5.5. What the convergence means, and the fine print on each price tag.
Deep dive07/26/2026
Tailwind CSS
Utility-first, JIT, design tokens, comparisons and when to use it.
Deep dive07/26/2026
Next.js + React
What each one is, how they work, when to use them and when not to.
Deep dive08/02/2026
Frameworks & deploy presets
Every framework in Vercel's import dropdown, grouped by what it actually solves.
Deep dive10/05/2026
DBMS: Database Management Systems
What a DBMS does, how it works inside (WAL, MVCC, indexes), every database family, main products compared and how to choose.
More posts coming soon: Git & GitHub, other databases, AWS, more languages to watch, API integration.