Not Replaced Yet

Anthropic just published Claude’s Constitution. An 84-page “soul doc” meant to shape how Claude thinks during training (this is NOT a system prompt). I read the whole thing and pulled out what actually matters: the priority stack (safety → ethics → Anthropic guidelines → helpfulness), the “1,000 users” policy lens, the no-white-lies honesty standard, the hard constraints (including the extremely explicit “don’t help destroy humanity” line), and the surprisingly candid section on Claude’s moral status + model welfare.
LINKS / SOURCES- Claude’s Constitution (official page): https://www.anthropic.com/constitution- Claude’s Constitution (84-page PDF): https://www-cdn.anthropic.com/cffd979fd050fbc0d8874b8c58b24cc10554e208/claudes-constitution_webPDF_26-01.26a.pdf- Anthropic announcement post (Jan 22, 2026): https://www.anthropic.com/news/claude-new-constitution- Older Constitution / Constitutional AI explainer (May 2023, updated Jan 2026): https://www.anthropic.com/news/claudes-constitution- Deprecation + model preservation commitments (Nov 2025): https://www.anthropic.com/research/deprecation-commitments- Amanda Askell's X: https://x.com/AmandaAskell
Chapters00:00 Intro (yes, it really says that)1:22 Our Vision for Claude's Character03:00 Section 1 — The Peculiar Position Admission03:59 Section 2 — The Four Priorities05:06 Section 3 — Why Helpfulness Matters05:23 Section 4 — Anthropic vs Operators vs Users (instruction hierarchy)06:34 Section 5 — Being Honest (no white lies / “epistemic cowardice”)07:28 Section 6 — The 1,000 Users Framework (responses as policy)08:20 Section 7 — Hard Constraints (bright lines)09:33 Section 8 — The Suspicion Clause (if it’s persuasive, get suspicious)10:14 Section 9 — Avoiding Concentrations of Power10:42 Section 10 — Being Broadly Ethical14:04 Section 11 — Being Broadly Safe18:38 Section 12 — Claude’s Nature21:03 Section 13 — Claude’s Well-Being23:54 Section 14 — A Final Word24:42 My Thoughts on Claude's Moral Consideration25:56 My Thoughts on the "Don't Help Destroy Humanity" Rules26:38 My Thoughts on Anthropic's justification for building the most dangerous technology ever27:27 Outro
Watch the video: https://youtu.be/7vsdvgDFiBk

Agents are ignoring your AI skill.
And if your whole workflow depends on that skill firing… that’s a problem.
In Vercel’s agent evals (testing modern Next.js APIs), the “skill” approach often didn’t get invoked at all so the pass rate didn’t improve. Then they tried a different approach that removes the decision point entirely and it hit a 100% pass rate.
In this video I break down:- Why skills fail to trigger reliably (and why telling the agent to use it is still brittle)- The exact pattern that made the agent consistently consult the right docs- How they avoided context bloat while keeping perfect accuracy- The bigger shift: from pre-training reasoning → retrieval reasoning
Chapters0:00 100% vs 53%0:08 Why skills are everywhere right now0:34 The eval results that changed my mind1:06 The real problem: outdated training data1:27 Skills vs “always-on” context (the key difference)2:17 Skills didn’t trigger in 56% of cases3:04 “Force the skill” helps… but it’s brittle3:52 Remove the decision point (the real fix)4:21 The results: 100% across build/lint/test5:00 How they avoided context bloat5:56 Practical recommendations6:29 Final thought + question for you
Watch the video: https://youtu.be/yEg-7sp9GLU

In the era of AI, agency is greater than intelligence.
I kept accidentally landing on the same post in my browser autocomplete. After seeing it over and over, it finally hit: in a world where you can rent genius-level reasoning for the cost of a subscription, intelligence isn’t the scarce resource anymore.
This video breaks down Andrej Karpathy’s “Agency is greater than Intelligence” idea, what agency actually means, and why this mental model matters right now if you want to stay ahead in the AI era.
Chapters0:00 Agency is greater than intelligence0:06 The “muscle memory” internet autopilot0:21 The autocomplete mistake that changed everything0:53 Too much AI signal to absorb1:32 The idea that kept repeating1:50 Karpathy + “Agency is greater than Intelligence”2:14 What “agency” actually means2:33 Why it matters right now2:36 Intelligence is being commodified2:58 Participant vs spectator3:07 Like / comment / subscribe
Watch the video: https://youtu.be/mkYSbHrnMFU

Claude Code’s latest updates aren't about one shiny feature. It’s the same architectural decision showing up in three places: MCP tool search, setup hooks, and hooks that can inject context at the moment of action. The theme is simple: load less at startup, surface more in the moment. This is Progressive Disclosure.
In this video, we break down how MCP tool search flips the old “load everything upfront” approach (and why that was burning huge chunks of your context window), how setup hooks let initialization run only when it’s actually needed, and why the new hook behavior (returning additional context pre-tool-use) is a big deal for anyone building agentic workflows.
If you’re wiring up lots of MCP servers, building custom tooling, or designing agent architectures, this update is worth understanding because it’s a pattern you’re going to see everywhere.
Chapters00:00 The “one decision” behind the update00:24 Welcome back to Not Replaced Yet00:40 Progressive disclosure explained (why it matters)01:19 MCP tool search: lazy-loading tools to save context02:26 Setup hooks: initialization at the right moment02:52 Hooks upgrade: injecting context right before actions03:41 The pattern across skills, tools, hooks, and setup04:09 Like + subscribe + questions in the comments
Watch the video: https://youtu.be/ozH00H98w6s

It’s 2026 and your second foundational AI upgrade is to stop using ChatGPT.
Like a chatbot, that is.
Chatting back and forth isn’t the unlock anymore. The unlock is getting AI to do things for you, not just tell you things. In this video, I break down the shift from chatbots → agents, show what “agentic” tools can actually do (web browsing, running code, creating files, interacting with apps), and run three real tasks across three agents so you can see the difference in outputs and workflows.
The big takeaway: stop thinking “What can I ask AI?” and start thinking “What can I delegate to AI?” Build the habit of reaching for agents first.
Chapters00:00 The 2026 AI upgrade: stop “chatting” with AI00:43 Chatbots vs Agents (what agents can do that chatbots can’t)01:14 A quick tour of agent options (ChatGPT, Manus, Perplexity, Claude Code, more)02:32 The underrated skill: context engineering02:51 Task #1: competitor research + 1-page report (ChatGPT Agent Mode)03:11 Task #2: repurpose a transcript into multi-platform content (Manus)03:39 Task #3: investigate a codebase + generate technical docs (Claude Code)04:10 Results: what each agent produced (and how it’s delivered)07:54 The mindset shift: “What can I delegate?”08:20 Wrap-up + what to do next
Watch the video: https://youtu.be/DryL2DarF4s

In this episode of Not Replaced Yet, I’m breaking down Claude Code v2.1.3. A surprisingly meaningful release with a real mental-model simplification (slash commands + skills are now one unified concept), plus a bunch of quality-of-life fixes that matter if you’re actually using this thing daily.
If you’re not using Claude Code yet, you’ll still get the vibe of how fast this tool is evolving. And if you are using it, these are the updates you’ll want to know about.
Links
Full Changelog: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md
Claude Code: https://code.claude.com/docs/en/quickstart
v2.1.0 video: https://www.youtube.com/watch?v=njAYrOOMenI
v2.1.2 video: https://www.youtube.com/watch?v=mFXR-JZUBOI
Chapters00:00 The “/doctor” joke + why it’s useful00:14 Intro + why v2.1.3 matters00:45 What’s covered in v2.1.301:02 Slash commands + skills merged (mental model simplification)01:54 Release channel toggle in /config (stable vs latest)02:11 Unreachable permission rules detection (+ fixes via /doctor)02:34 VS Code permission destination selector (project/all/team/session)03:06 Plan file reset fix after /clear03:27 Skill duplicate detection fix (XFAT edge case)03:51 Background task count mismatch fix04:06 Sub-agent model fixes (correct model during compaction + web search)04:33 Trust dialog fix from home directory (hooks properly enable)04:57 Terminal rendering stability improvements05:08 Slash command description truncation (2-line autocomplete)05:29 Hook timeout extended: 60s → 10 minutes05:51 Wrap-up + January 2026 pace is wild
Watch the video: https://youtu.be/nRa77FXf3ow

In this episode of Not Replaced Yet, I break down the latest Claude Code updates (v2.1.2). The Anthropic team is shipping at a wild pace.
If you’re already using Claude Code, these updates remove real friction from everyday dev workflows. And if you aren’t using Claude Code yet, get on board!
We cover everything from image source metadata and clickable file paths, to massive improvements around large outputs, permissions, Windows installs, and — most importantly — a critical security fix you do not want to skip.
Chapters00:00 Claude Code is shipping fast00:50 What’s new in v2.1.201:03 Image source path metadata01:35 Clickable file paths in output02:01 Shift+Tab auto-accept in Plan Mode02:18 Bash output saved to disk (no more truncation)02:55 Large tool output persistence03:16 Permission system improvements03:43 Unified Plugins + MCP view04:10 Windows install via Winget04:24 ⚠️ Command injection security fix05:15 Memory leak fix05:34 Binary files loading bug fixed05:56 Image paste over SSH improvements06:11 macOS Meta key setup guidance06:27 Socket file crash fix06:43 Final thoughts on Claude Code’s pace
Watch the video: https://youtu.be/mFXR-JZUBOI

In the time it took you to click this video, Claude Code shipped another update 😅
Today I’m breaking down the major Claude Code 2.1.0 release (and the quick follow-up patch), including the biggest workflow wins: automatic skill hot-reload, forked skill execution, wildcard Bash permissions, hooks everywhere, real-time thinking display, and more.
If you’re using Claude Code already, these are the changes you’ll feel immediately. If you’re not using it yet… get started!
Chapters00:00 Claude Code shipped again (lol)00:13 Welcome to Not Replaced Yet00:44 Claude Code 2.1.0 overview (+ quick patch context)00:55 Automatic skill hot reload (no more restarts)01:14 Forked skill execution (context: fork)01:48 Agent type specification in skills02:14 Wildcard Bash permissions (* anywhere)02:56 Hooks in agent/skill/command front matter03:29 Hooks “once: true”04:04 MCP list_changed notifications04:51 Real-time thinking display (Ctrl/Cmd+O transcript mode)05:24 Plugin hook support (global enforcement)05:55 /plan shortcut06:12 Pre-tool hooks can modify tool inputs (updatedInput)06:49 Background task management (Ctrl/Cmd+B)07:13 File read token limit override env var07:53 Slash command autocomplete anywhere08:16 Skills show up in the slash menu08:37 Better command search prioritization09:10 Reduced permission prompts for chained Bash commands09:45 Wrap-up
Watch the video: https://youtu.be/njAYrOOMenI

60% of the time, vibe coding works every time 😏
Welcome back to Not Replaced Yet. This is Part 2 of vibing an app into existence. In this episode I jump back into the NRY website build, make quick UI tweaks, add a smooth entrance animation, test the admin/dashboard flow (Clerk auth + Supabase-backed content), and then go from “it works locally” to a real production deploy on Vercel… including hitting (and fixing) a Next.js security/build issue mid-flight.
Big theme of this series: vibe coding is great when the stakes are LOW!
Chapters0:00 “60% of the time…” + intro0:22 Part 2 setup + recap of Part 11:04 The rule of vibe coding: keep stakes LOW1:45 /init + cloud.mmd context trick (why it helps)2:53 Homepage tweaks (remove background, stack headline, tighten subhead)4:54 Adding a soft fade/float-in text animation6:39 Fixing animation choppiness (timing + load interaction)9:53 Blog + Clerk auth test (finding sign-in route)12:50 Creating a blog post via ChatGPT + publishing workflow14:54 Testing “Signal” + how links behave17:20 Building the Resources page + featured items20:23 Creating a new private GitHub repo + first push23:15 Importing into Vercel + env vars26:56 Vercel build failure (Next.js vulnerability)28:49 Security review callout + updating Next.js33:10 Why you should be talking to your computer (Whisper Flow)35:45 Build succeeds + auto-deploy to production37:36 Live site tour (Home / Blog / Signal / Resources)38:09 Wrap-up + what’s next
Watch the video: https://youtu.be/vreNa2rWw_E

In this first episode of a two-parter of Not Replaced Yet, I go off-the-cuff with a live vibe code session for the NRY website. We start from zero in Claude Code and let the agent cook.
Along the way, I show the exact workflow I’m using in 2026:• Plan Mode for specs and clarifying questions• Design inspiration loops (Dribbble + Landing Love)• Copy-paste UI acceleration with 21st.dev prompts• Supabase + MCP for database setup and migrations• Sub-agents running in parallel with fresh context windows• A simple Claude Code hook that plays an ElevenLabs voice when a task finishes• Real debugging when things inevitably break (React 19 + hydration issues + styling fixes)
This is low-stakes build mode — don’t do this for serious client work. But for personal projects and fast iteration, it’s insane.
Part 2 is coming next, where we finish the remaining UI polish, fix the last issues, and ship the site.
Chapters0:00 Intro and what we’re building0:20 What “vibe coding” actually means (and when it’s OK)1:24 Why your own website always gets neglected2:56 Plan Mode setup and the initial prompt5:10 Finding design inspiration on Dribbble + Landing Love10:19 Using 21st.dev prompts to speed up UI14:08 Supabase setup and adding MCP20:39 Context windows and why quality degrades near the limit22:36 How the Claude Code creator runs multiple agents in parallel29:12 Spinning up sub-agents to do research without burning context32:18 Hooks that run when an agent finishes34:32 Adding an ElevenLabs audio cue for “task complete”45:26 Clerk auth setup (admin-only access)54:48 Build completes and what got generated55:42 /compact and starting the next phase cleanly58:45 Applying the Supabase schema/migration via MCP1:00:29 Running the app locally and wiring env vars1:04:17 React 19 issue and a full shader rewrite in vanilla Three.js1:07:44 First look at the site and quick UI checks1:10:57 Fixing the invisible headline (opacity issue) and remaining hydration error1:12:16 Wrap-up and Part 2 teaser
Watch the video: https://youtu.be/jOop4g8_ZZM