Videos

Creator codenameposhan shares "Obsidian Skills," a Claude Code plugin marketplace package built by Steph Ango (Obsidian's CEO) that teaches Claude Code Obsidian's native format — proper notes, real file connections, canvas diagrams, vault-wide terminal control, and clean markdown from messy web pages.

Quick explainer on SQL injection attacks for non-technical builders: covers what SQL is, how malicious user input can be injected into database queries as executable commands, and two practical defenses. Part of a daily series teaching technical security terms to vibe coders.

Step-by-step walkthrough of using Claude Cowork's Apify connector + Scheduled Tasks to automatically pull Instagram metrics every Sunday and generate a dashboard summary. Creator shows the exact prompt used and the resulting dashboard output.

Skit-style system design explainer showing why polling a server every second fails at scale (5M requests/sec). Covers the progression from naive polling → caching → WebSockets → Pub-Sub → conflation, using Robinhood's live stock price feed as the example.

Creator aimn.dev demonstrates a Claude skill called "ponytail" designed to fight AI over-engineering in vibe-coded projects. Before writing a single line of code, the skill forces Claude to ask: does this need to exist? Can the standard library do it? Is there already a one-liner? Only if nothing simpler works does it write code — and only the minimum. Claims: 90% less code, faster review, cheaper token costs.

arjay_the_dev presents the senior engineer approach to testing: targeting 100% line and branch coverage before opening a PR, combining unit tests with integration tests, and designing code for testability from day one using dependency injection, pure functions, and separation of concerns.

davebrownbrand argues the real Claude upgrade isn't a better prompt — it's connecting Claude to the right tools. He walks through three plugins: Anthropic's Productivity plugin (connects Gmail, Calendar, Slack), a Brand Voice plugin (writes copy in your brand's tone), and a Marketing plugin (advises on landing pages, SEO, re-engagement campaigns).

Claude Code sessions degrade over time as context fills with broken reasoning and failed attempts. /compact alone can carry that forward. A handoff.md file captures the goal, current state, files in progress, changes made, failed attempts, and next steps so a fresh /clear session starts informed.

Demo of Clips, a free open-source Chrome extension for recording screen + audio + browser logs/network requests in one link. Unlike Loom, Clips outputs metadata APIs so agents can directly pull transcripts, video snapshots, and logs from the URL alone — no MCP server needed. Useful for reporting bugs to agents with full context.

Creator softgirlnocode argues that plugging AI into individual apps is the "last mile problem." The fix is an AI Operating System: a local repository holding all your business context (contacts, brand assets, ongoing projects, quarterly plans, software data) with AI connected directly as a co-pilot. Example: a new hire onboarding doc is generated automatically because the OS already has all contacts, references, and project context.

Hayden Schmitty walks through five common performance and UX issues that show up in AI-built apps: missing button tooltips, no optimistic UI rendering, unpaginated backend queries, N+1 database query patterns, and synchronous operations. Each issue is framed as a quick vibe-codable fix.

techwithnt breaks down Amazon security engineer Eric Brandwine's June 2026 argument that Human-in-the-Loop is no longer the gold standard for AI safety. When humans approve endless agent actions, they become 'exhausted approval buttons' — a phenomenon called normalization of deviance. The solution: shift from human approval to end-to-end agent accountability, where every agent has its own identity, permission boundary, and activity trail.

itsthatlady.dev explains why pushing directly to main is always wrong — even solo — and walks through the four-branch strategy: feature branches for new work, dev as the active integration branch, staging as a production mirror for pre-launch testing, and main as the sacred production branch protected by branch protection rules and required reviews.

itsmariahbrunner covers Anthropic's Claude Tag feature, which adds Claude to a team's Slack workspace as a first-class member with its own account and admin-set permissions. Tag it in any channel and it reads the full thread, understands context, and takes action — without impersonating any individual user's login.

steve8708 shares three daily-use Claude Code skills: Agent Watchdog (have one agent monitor and fix another agent's work), Planarbiter (two agents make separate plans, negotiate differences, and produce a final plan), and 'Read The Damn Docs' (forces agents to look up actual documentation instead of guessing at APIs). He also describes a workflow combining all three: Planarbiter → Claude Code writes the plan → Codex executes → Claude Code watchdogs.

buildwitholi explains the two types of agentic loops: deterministic loops (measurable targets Claude checks itself — e.g. page load under 50ms, zero errors, 100% test coverage) and LLM-as-judge loops (no clean metric — Claude evaluates its own work against a rubric, e.g. 'refactor until clean' or 'update docs until they reflect the codebase'). The rule: convert to deterministic whenever possible; only use LLM-as-judge when genuine judgment is required.

haydenschmitty lists five backend and frontend performance issues that Claude typically generates without fixing: uncompressed JSON responses, single-row database inserts (should be batched), single dependency bottlenecks detectable via roundtrip latency audit, no optimistic UI updates (UI waits on every backend response), and server-side HTML rebuilt for every visitor (should be cached/static).

designmotionhq covers five UX engineering rules for drag-and-drop file uploads: drop zones must give visual feedback (border, glow, copy shift) before the drop; progress bars should show percent and time remaining — not a spinner; failed uploads must offer inline retry without losing the file; thumbnails/type/size should confirm receipt; and multi-file uploads need independent progress and retry per file.

A single CLAUDE.md file based on Andrej Karpathy's notes has 170,000 GitHub stars. Drop it in your project root and Claude Code follows four constraints: think before coding, keep it simple, only change what was asked, and loop until done. Creator sends the GitHub repo and install guide via DM (comment "RULES").

timosterhus draws a sharp line between vibe coding (open-ended requests like 'make me an app') and agentic engineering (following software engineering best practices with an agent). Agentic engineering means: write a detailed spec, decompose it into scoped work packets, implement, test, and keep the whole system coherent throughout. Code quality difference is described as 'night and day.'

raycfu demonstrates a two-agent Claude Code loop: a Builder agent that writes and fixes code, and a Checker agent that runs tests, type checks, and linting. An orchestrator chains them in a loop — Builder implements, Checker validates, failures get sent back to Builder until all green. Crucially, two stop rules prevent infinite loops and token burn: max 5 cycles, and immediate stop if a fix breaks something previously passing.

pikacodes interviews Neel Sundaresan, creator of IBM's enterprise coding agent "Bob," on deploying AI coding tools safely at scale. Key themes: model evaluation for compliance before adoption; giving developers control via custom modes, RAG, and tool-calling; designing against alert fatigue so engineers don't disable security scanners; and framing AI as an enhancer (not replacement) for engineers — 60% of software work is modernization/migration, not new code.

creativelyange covers Claude Code's new live artifacts feature — at the end of a coding session, Claude Code generates an artifact that summarizes all the code, PRs, and project progress. The artifact is shareable with teammates as a project overview.

Creator codeyourai demonstrates the Anthropic tip that "HTML is the new markdown". When learning a new tool (Remotion), instead of reading a wall of text from Claude, ask it to convert the explanation to HTML — you get a structured page with sections, code examples, and install steps.

haydenschmitty shares a free Claude skill that audits a codebase against 300+ common Claude-generated mistakes across three pillars: UI/UX design, performance, and security. The skill runs for several hours uninterrupted, so you can start it before bed and wake up to an improved codebase.

jonathan.lewell walks through a workflow to capture YouTube videos into an Obsidian knowledge base using the Obsidian Web Clipper extension, which pulls in the video transcript and metadata. Then inside Claude Code (co-work mode), he queries the note by path to extract atomic insights and writes them back to the top of the note above the transcript.

qendresahhoti explains how to build a personal Claude operating system: one root folder with domain subfolders (content, business, etc.), a me.md describing who you are and how you work, and a CLAUDE.md (lot.md) that maps what lives where. Claude reads the map each session, loads the right files into context, and answers from your actual data — nothing is stored, everything is re-read.

Create one master context file (agents.md) and symlink CLAUDE.md and other AI tool config files to it so all tools read from a single source of truth. Eliminates duplicate context maintenance when using Claude Code + Codex + other AI tools. DM-gated: comment 'MIKA' for symlink command and guide.

A free Claude Skill called 'Last 30 Days' searches Reddit, X, YouTube, Polymarket, and the general internet but filters results to only the last 30 days, then feeds that into Claude before it answers. Solves the problem of AI confidently citing stale (year+ old) information as current. DM-gated: comment 'last30' for install instructions + usage tips.

Creator explains why serious Claude users work in VS Code instead of the Claude app: (1) Not locked to one model — Claude and Codex both available, swap to whichever is better that week; (2) Marketplace with thousands of free extensions to customize the workspace; (3) Every AI tool she uses connects to VS Code so everything lives in one place instead of bouncing between apps. DM-gated: comment 'vs' for setup guide and extension list.

Part 1 of Claude Code Summer School. Covers the 4-step workflow: Explore, Plan, Code, Commit. Key insight: use Plan Mode (Shift+Tab until 'plan mode' shows) so Claude reads/thinks before touching any files. Also covers /init to explore large/unfamiliar codebases. Demonstrates with webcam gesture control project.

Follow-up to the loop engineering intro video (also by @derrickfromzero). Demonstrates building a real loop using Claude.ai's scheduled tasks feature. Pattern: Task 1 'Sourcer' runs hourly with a prompt to generate business ideas and write them to a Notion database. Task 2 'Checker' also runs hourly, reads the Notion database, and validates each idea. The two tasks alternate, creating a self-refining feedback loop. Mentions adding a third 'check the checker' task for more rigor.

Microsoft paper 'SkillOpt' shows that a second AI can watch an agent complete tasks, identify where it failed vs. succeeded, and propose tiny edits to the agent's skills doc (the SOP/system prompt) to improve performance — without touching model weights. GPT-4.5 spreadsheet accuracy went from 41% to 80%; document tasks from 33% to 72%, with gains coming from just 1–4 edits. Won across 52 model/task combinations. Limitation: only works when there's a clear right/wrong answer to grade against (not creative/strategic tasks).

@steve8708 built two complementary Claude Code skills: (1) /visual-plan — generates a plan as MDX with visual components: diagrams, interactive API specs, schema changes, annotated code, pan/zoomable wireframes. Review the wireframe, comment/iterate, then let the agent implement. (2) /visual-recap — after agent finishes, generates a visual recap of everything it did (same format: wireframes, diffs, annotated code). Useful for reviewing a PR or another dev's work. Creator's insight: humans are now reasoning at the plan level, and as long as the plan is right, agents are increasingly reliable at executing it — similar to trusting a compiler.

2-minute clip from a Boris Cherny (Anthropic staff, Claude Code creator) conference talk. This clip covers the intro: what Claude Code is (fully agentic, works with all IDEs, builds features/functions/fixes bugs — not line completion), and a live install demo. The full 24-minute talk reportedly covers 40 real prompting techniques from the person who built it. This clip is the teaser; the full video is linked/referenced by @aiforgez.

Short reaction clip: creator claims to have found a 'loophole' to access Claude Fable 5 through AWS Bedrock Playground during what appears to be a regional/rollout restriction. Low production quality, unclear if the method still works or was ever valid.

Conceptual intro to loop engineering vs. prompt engineering. Prompt engineering = one-shot: give context, get answer. Loop engineering = the AI gets a 'second opinion' (another model/pass) that must verify the result before it's returned. The verification loop prevents low-quality answers from passing through. Uses doctor analogy: first opinion → second opinion must agree before diagnosis is given.

Short demo clip (credited to @om_patel5): a Kanban board UI for Claude Code where dragging a task card into 'In Progress' immediately triggers a Claude Code agent session on that task. Supports running multiple agent sessions simultaneously, tracked from a single board — manage agents like a project manager.

Creator outlines 3 progression levels for Claude Code: Level 1 — using VS Code extension, not knowing how to give context, stuck in correction cycles. Level 2 — CLI user who knows CLAUDE.md, their stack, how to debug, when to use Plan Mode, and how to direct vs. correct. Level 3 — expert at context management, running multiple agents across worktrees simultaneously, using /remote (builds even when not at the computer — 'watching interns do the work').

Token management advice using the 'backpack' analogy. 3 steps: (1) CLAUDE.md — put rules + brand voice in one file Claude reads automatically, so you stop re-explaining yourself every message (in Claude.ai: add to Project; in Claude Code: global MD); (2) Context folder — put heavy reference docs (brand docs, past projects) here and Claude only pulls them when needed, not for every prompt; (3) Fresh chat per task — don't drag a full day of context into a 5-minute task. DM-gated: comment 'SNAIL' for full token-saving guide.

Second video from @nelsontalksmoney (ex-Shopify, Solarium startup). Same core workflow as his earlier video but adds an important bonus insight: use the most powerful model (Opus 4.8) for planning/spec writing, then use a lighter model for the actual coding. Full workflow: (1) Tech spec in Claude Desktop + Opus 4.8; (2) Hook Claude Code to GitHub, convert spec to Epic → Issues → Sub-Issues (traceable, scoped, reviewable); (3) Execute issue by issue with a lighter model. Framing: the problem isn't the model version — it's skipping the spec step entirely.

itsmariahbrunner offers a safer alternative to enabling full bypass permissions in Claude Code: build an explicit allow list of safe commands (npm run, git status, reading files) that Claude runs instantly without asking, and a deny list of dangerous commands (rm -rf, git push --force, sudo) that Claude explicitly blocks and cannot override. The result is the same speed as bypass mode without the risk of destructive commands running unchecked.

Creator demonstrates why LLM outputs are unpredictable using a book tracker app where the model tagged Dracula as 'cuddly, family friendly, kittens.' Introduces evaluation frameworks as the solution: define metrics, auto-grade outputs with LLMs, analyze trends over time, and synthesize test samples automatically.

Setup pattern using CLAUDE.md + handoff.md + two hooks for session continuity: (1) Run /init to create CLAUDE.md (codebase overview, loads every session). (2) Prompt Claude to create handoff.md with summary of last session + what to work on next. (3) Start hook: reads handoff.md automatically at session start alongside CLAUDE.md. (4) Stop hook: auto-updates CLAUDE.md with big changes and handoff.md with what was worked on + what's next. Result: Claude always knows the project context and next task without you having to remember between sessions.

Ex-Shopify engineer (now AI startup founder) explains his 3-step production Claude Code workflow: (1) Write detailed tech specs in Claude Desktop with Opus 4.8 — full markdown docs covering edge cases, data flow, UI/UX before touching Claude Code; (2) Connect Claude Code to GitHub, have it turn the spec into an Epic → Issues → Sub-Issues, creating a paper trail with clear scope and milestones per ticket; (3) Execute issue by issue — small, reviewable chunks, not one giant black box. If something breaks, you know exactly which issue caused it. One-shot demo approach fails on real production codebases.

itsmariahbrunner covers Claude Code's Dynamic Workflows feature, which lets you tell Claude to create a workflow and it spins up tens to hundreds of sub-agents all working in parallel. A second wave of agents then reviews the first wave's work for errors. Example: a developer rebuilt 750,000 lines of code in 11 days using hundreds of parallel agents — a job that previously took months.

Vercel's Agent Browsers gives Claude a dedicated browser session using an accessibility-file approach instead of screenshots. Compared to Claude for Chrome or Playwright MCP (which send full screenshots and large tool definitions), Agent Browsers provides smart element references — consuming 15x fewer tokens per interaction.

Creator built a Claude skill called "Lock It In" that solves the problem of vague skill exports. When you finish a workflow you love, instead of asking Claude to "turn this into a skill" (which produces vague summaries), you type "Lock it in." The skill: (1) reads entire conversation to find what worked vs what you rejected, (2) asks 2-3 clarification questions you didn't think to answer, (3) writes the skill in proper format with examples + edge cases, (4) self-tests with fake input before handing off. DM-gated: comment "lock."

Creator explains the critical difference between authentication (sign in/log in) and authorization (what you're allowed to do once inside). AI tools handle auth setup fine but routinely forget to add authorization checks, leaving users able to access other people's data by simply changing an ID in the URL.

Teaser clip explaining what an AI harness is — the environment built around an LLM to make it more effective at specific tasks. References Claude Code and Codex as harness-based systems. Full video available separately (see creator's profile). DM-gated link: comment anything on their post.

Creator tested dozens of tools and identified 4 essential Claude Code add-ons: (1) Context7 MCP — feeds current library docs on demand so Claude doesn't reference outdated APIs; (2) Sequential Thinking MCP — adds a reasoning layer so Claude pauses on hard problems instead of charging through; (3) Superpowers MCP — forces Claude to back up work before claiming done, enables serious engineering workflow; (4) Plan Mode (built-in) — forces research + your approval before touching any file. DM-gated: comment 'BUILD' for install guide.

Three underused Claude Code features: (1) Sub-agents — instead of researching in your main session (burns 10% context per task), tell Claude to 'spin up a sub agent to research this' — a separate instance does the work and returns only the summary, using just 0.5% of main context; (2) /status shows a color-coded real-time context meter (green=good, orange=slow down, red=clear out); /context breaks down exactly what's consuming space (system prompt, MCP servers, conversation history); (3) Hooks for auto-skill routing — a small script watches what you type and automatically loads the right skill (research, script-writing, product building, competitor analysis) without you specifying which one.

Creator walks through 3 quick security checks for AI-generated code: (1) use env variables for API keys — never hardcode them; (2) use battle-tested auth providers (Supabase, Clerk, NextAuth) instead of AI-invented auth; (3) prompt AI as a hacker to find and fix its own vulnerabilities before you ship.

Team exercise where everyone blocks 1 hour to have AI interview them via prompted back-and-forth, building reusable context docs covering personal goals, team goals, family, friends, and vendors. Uses voice dictation (Wispr Flow) for efficient input. Context docs are then reused in future AI sessions to improve output quality.

References Anthropic's 33-page guide on Claude Skills. A Skill is a reusable instruction pack built once that Claude executes consistently every time. On Team/Enterprise plans, one person can deploy a skill across the entire company. Creator's career advice: build a skill for a repeated team workflow, bring it to your manager, keep a running list of skills you've shipped, and use that list as promotion/raise evidence. DM-gated: comment 'SKILLS' for 10 ready-to-use templates.

Creator built a slash command in Claude Code that's a PRD builder designed to double as the prompt for the coding agent. Key design choices: (1) Claude designs the testing suite and success criteria BEFORE implementing features; (2) runs unit/integration tests and git commits after every significant feature; (3) re-tests past features after new ones are added; (4) leaves space in the PRD doc for the agent to write notes to itself for later sessions. Drastically reduced back-and-forth messages needed to ship a project.