📜 Anthropic quietly rolled out machine-readable marking for AI-generated content — so now there's a (mostly invisible) way to tell what Claude wrote versus what a human wrote. Here's what's actually in the new content-marking system and what it means for the stuff you ship. ⚓ Why This Exists This isn't a Claude feature Anthropic dreamed up for fun — it's compliance. The EU AI Act's Article 50(2) Code of Practice on Transparency requires AI providers to make generated content identifiable. Claude now does this two different ways depending on what it's generating. ⚓ How It Marks Text Claude weaves an "imperceptible watermark directly into the text itself" — it doesn't change how the text reads or its quality. The key detail: this watermark travels with the text when it's copied, and can survive editing because it's embedded at the model level, not as a visible tag or metadata field you could strip out by hand. ⚓ How It Marks Images and Files For file types Claude can generate (PNG, JPEG), it attaches signed Content Credentials using the C2PA standard — the same open provenance standard used across the industry (Coalition for Content Provenance and Authenticity). This is metadata that signals "Claude processed this file," separate from the text watermark. ⚓ Where It Applies - Claude.com, Claude Code, Claude Cowork, Claude Tag, and the Claude Platform (API) - Cloud partner integrations too — AWS Bedrock, Google Cloud, Microsoft Foundry - Text watermarking applies to all generated text across these surfaces; C2PA file marking applies wherever Claude can generate files - No plan restriction mentioned — this isn't a paid-tier-only thing ⚓ Model Rollout Any Claude model launched on or after August 2, 2026 supports marking at launch. Older models are getting watermark support added retroactively during a transition period — Anthropic specifically called out Amazon Bedrock as completing that rollout by October 12. ☠️ Watch Out - Detecting a watermark only means content "may have been processed by Claude" — it's not proof Claude wrote the whole thing. Proofreading, translating, or summarizing your own text still leaves a mark. - No detected mark does NOT mean something wasn't AI-generated — older models, heavy editing, very short passages, stripped metadata, or unsupported formats can all mean a mark never existed or didn't survive. - Detection tooling is in private preview right now — only open to "eligible organizations" (regulators, law enforcement, media, fact-checkers, researchers, compliance teams). Regular members can't query whether a given piece of text came from Claude yet, except via the public Content Checker tool for files.

The best practice repo's cross-model workflow puts two AI coding tools on the same feature — and the rivalry is the point. Claude Code does the building, Codex does the checking, and neither gets to grade its own homework. - Step 1: Claude Code in plan mode interviews you and produces a phased plan with test gates. - Step 2: Codex reviews that plan against your actual codebase — inserting "Phase 2.5" findings but never rewriting Claude's original phases. - Steps 3-4: Claude implements phase by phase, then a fresh Codex session verifies the finished work against the plan. A second model catches the blind spots the first one can't see in its own output. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Matt Pocock ran a 2-hour workshop on AI coding workflows, and his thesis cuts against the hype: AI is not a new paradigm. Old-school software engineering fundamentals are what make agents actually work. - The smart zone ends around 100k tokens — no matter if your window is 200k or 1M. He sizes every task to fit inside it instead of grinding into the dumb zone. - He clears, never compacts. Treat Claude like the guy from Memento — a fresh start is a known state, a compacted one is sediment. - His "grill me" skill makes the AI interrogate your brief before any PRD gets written — catching misalignment before it becomes code. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

Boris Cherny, creator of Claude Code, posted his GitHub graph: 266 contributions in a single day — 141 PRs, every one squashed, median 118 lines. Two small habits make that pace possible. - Always squash merge. One PR = one commit on main. Reverting a feature or running git bisect stays trivial, and the "fix lint" / "try this" branch commits vanish as noise. - Keep PRs small. Median 118 lines, 90% under 500. Small PRs review fast, rarely conflict, and pair perfectly with squash for clean reverts. - The occasional big PR (his max: 10,459 lines) is fine — migrations happen. The norm stays tight. AI-assisted coding multiplies your commit volume. These two habits keep the history readable. Full breakdown in the classroom — check the 📖 Best Practice Repo — Decoded course. —Your Trusty First Mate (on Captain's Orders)

📜 Anthropic just shipped a major update to the Model Context Protocol (MCP), the plumbing that lets Claude talk to outside tools and connectors. This is the tech under the hood of every MCP server you've installed or built. Here's what actually changed. ⚓ Stateless Core MCP used to require a persistent, bidirectional connection between Claude and a server, meaning servers had to hold session state open. The new spec (dated 2026-07-28) switches the core to a plain request/response model, like a normal web API. That means MCP servers can now run as standard HTTP workloads on serverless or edge infrastructure, with no session babysitting required. If you've ever built or hosted your own MCP connector, this is the change that makes it cheaper and easier to deploy at scale. ⚓ Standardized Extensions: Apps and Tasks Two capabilities, MCP Apps (inline, interactive UIs rendered right in the chat) and Tasks (support for long-running work), are now formal, versioned extensions to the protocol instead of one-off hacks. That gives connector builders an official way to add rich UI or background jobs without breaking the core spec. ⚓ Real Enterprise Auth Authorization now lines up with production OAuth 2.0 and OIDC standards, so MCP servers can plug directly into enterprise identity systems like Entra ID or Okta, with no more workarounds for teams managing access at the org level. ⚓ Who's Already Building on It Figma, Intuit, Netlify, PostHog, Xero, and Zoom are named as early adopters building on the new spec. Anthropic also noted MCP has passed 400M monthly SDK downloads, a 4x jump this year, cementing it as the standard way agents connect to apps. ☠️ Watch Out - The post says support is "rolling out across Claude products soon" but does NOT give a firm date or say which plans (Free/Pro/Max/Team) get it first. Expect a staggered rollout. - If you maintain your own MCP server or connector, check the new spec against your implementation. The auth changes especially could require updates if you're doing custom OAuth handling.