Claude Code Intel
10 videos · 7 channels · 44+ actionable ideas. Enter hub password to unlock.
That's not right. Try again.
Claude Code Research
Claude Code Intel
May 2026 · 10 videos watched across 7 channels
10
Videos
7
Channels
44+
Tips
Most Mentioned (3+ videos)
Highest confidence- All 10CLAUDE.md file, Create one in every project. Claude reads it every session. Keep under 200 lines; route to sub-files for longer content.
- 9 videosPlan Mode (Shift+Tab twice), Claude maps every step before touching files. Review and iterate before committing. Never skip this.
- 8 videos/clear between tasks, Prevents context rot degrading output quality on the next task.
- 7 videosultra think keyword, Type these two words before any hard prompt. Allocates 32K token thinking budget. Zero setup.
- 6 videosSub-agents on Haiku, Run bulk reading and scraping sub-agents on Haiku. Keep Opus or Sonnet on the main synthesis thread only.
- 6 videos/context command, Shows exact token breakdown by source. Run before you hit limits. Disconnect unused MCP servers; each loads ~18,000 tokens per message.
- 5 videosDisconnect unused MCP servers, Each connected server can eat ~18,000 tokens per message even if you never use it that session.
- 4 videosWAT Framework, Nate Herk's three-layer structure: Workflows (markdown SOPs) + Agent (Claude Code) + Tools (Python scripts). Everything else builds on this.
- 4 videosContext Mode plugin INSTALLED, Sessions run 3 hours instead of 30 minutes. 315KB tool output compressed to 5KB. SQLite session log survives compaction.
Skills
12 found| Skill | What it does | Install | Note |
|---|---|---|---|
| Skill CreatorHOT | Describe a skill; Claude builds, tests, and packages the .md file | /plugin install skill-creator | Mentioned 3x |
| GSD (Get Stuff Done)HOT | Spawns fresh sub-agents per task; scope detection catches dropped requirements | Nate Herk community | Fights context rot at infrastructure level |
| SuperpowersHOT | Forces plan-first; built-in TDD loop; two-stage self-review | /plugin install in Claude Desktop | Already in your stack |
| Context ModeINSTALLED | Sandboxes tool output; 315KB to 5KB compression; sessions run 3h | Installed this session | Most impactful single install |
| ClaudeMemINSTALLED | SQLite + vector memory across sessions; auto-generates folder CLAUDE.md files; 10x token savings | Already in your harness | Handles cross-session persistence |
| /ultra review | Fleet of parallel cloud reviewers; reproduces bugs independently before listing | Built into Claude Code v2.1.86+ | $5-20/run; 3 free on Pro/Max |
| Karpathy SkillsFREE | 4-principle CLAUDE.md: think first, build simplest, surgical changes, don't improve what wasn't asked | One-file install from GitHub | Add to every project CLAUDE.md |
| Open DesignFREE | 72 brand-grade design systems (Stripe, Airbnb, Notion, Linear) baked in | git clone + install globally | Best for web and visual builds |
| /insights | HTML report on last 30 days of your Claude usage, graphs, patterns, anti-patterns | Built into Claude Code | Run monthly. Guaranteed surprises. |
| Morning Coffee | Reads calendar + goals + project status; proposes full day plan; blocks calendar | Build it yourself (Nate Herk template) | From Executive Assistant video |
| /btw | Ask Claude a side question mid-task without interrupting the main task | Built into Claude Code | Non-obvious built-in |
| /cavemanFREE | Cuts output token volume significantly; useful on lower-tier plans | Open-source community skill |
Workflows & Patterns
9 documented1
WAT Framework (Nate Herk's signature)
- Layer 1, Workflows: plain-English markdown SOPs in /workflows/ folder
- Layer 2, Agent: Claude Code reads SOPs, decides how, handles failures
- Layer 3, Tools: Python scripts in /tools/ that execute exactly one action
- Self-healing: errors trigger research + fix + SOP update automatically
- Deploy Workflows + Tools to production (Modal, trigger.dev); agent is development-only
- WAT CLAUDE.md template available free at github.com/nateherkai/AIS-OS
2
Executive Assistant 4-Phase Build
- Phase 1: CLAUDE.md + sub-files (me.md, work.md, team.md, priorities.md)
- Phase 2: Onboarding interview stores your preferences in structured context files
- Phase 3: Skills for each tool (Perplexity research, project management MCP)
- Phase 4: Migrate all custom GPTs into skills; say "remember I always prefer X"
3
Plan Mode to Bypass Permissions (two-phase pattern)
- Step 1: Shift+Tab twice into Plan Mode; Claude asks questions and produces plan with cost estimate
- Step 2: Review, add brand assets via @ tagging, redirect bad decisions
- Step 3: Switch to Bypass Permissions; Claude builds without interruption
- Never use Bypass without prior planning phase
4
Self-Improving Workflow Loop
- CLAUDE.md instruction: "When you hit an error, research the fix, update the tool, update the SOP"
- After each run + feedback, Claude edits its own workflow markdown files
- Compound improvement: the more you run it, the better it gets
5
Competitor Research + Branded PDF
- One-time setup: create a business_profile.json with your business context
- Trigger: "Do competitor analysis for my business"
- Claude auto-detects workflow, scrapes competitors (Firecrawl), generates branded PDF
- Cost: ~$1.50 first run, ~$0.50 subsequent (data cached in /competitors/ folder)
6
Routines (Cloud Automation via /schedule)
- /schedule + plain-English description; Claude writes the system prompt for you
- Two trigger types: time-based (cron) or event-based (GitHub events, custom webhook POST)
- Agent-on-agent review: Routine A creates PR, Routine B reviews it automatically
- Intervene in any running session from your phone via claude.ai web
7
Parallel Worktree Sessions
- claude --worktree <feature-name> creates isolated Git branch workspace
- Run 3-5 simultaneous sessions on the same project with no file conflicts
- Sound notification hook tells you when any session finishes
8
Screenshot Self-Check Loop (web builds)
- Bake into to-do list: build step → screenshot → analyze → revise
- Three passes before calling V1 done
- Chrome DevTools loop: Claude opens browser, interacts, checks console errors
9
YouTube Shorts Factory with Remotion
- npx create-video@latest installs Remotion (free, open source)
- One Claude Code prompt builds the full template with animations
- Store all video content in a JSON data file; template stays fixed
- Batch render script processes all JSON entries in one pass
- Always validate topic virality score (aim for 92%+) before batch production
Plugins & Extensions
8 found| Plugin | What it does | Install |
|---|---|---|
| Context ModeINSTALLED | Sandboxes tool output; 315KB to 5KB compression; 3-hour sessions; SQLite session log | Installed this session via mksglu/context-mode |
| ClaudeMemINSTALLED | SQLite + vector memory; auto-generates folder CLAUDE.md files; 10x token savings | Already in your harness (claude-mem skill) |
| SuperpowersINSTALLED | Forces plan-first; TDD loop; two-stage self-review; auto-promotes branches | Already in your stack |
| Everything Claude CodeFREE | 48 agents, 182 skills, 68 commands; cross-session memory; CVE scanning | /plugin install everything-claude-code |
| Open DesignFREE | 72 brand design systems; Stripe/Airbnb/Notion aesthetics on demand | git clone nexu-io/open-design + install globally |
| GSDFREE | Scope detection; fresh sub-agents per task; security enforcement; quality gates | Nate Herk AIS-OS GitHub repo |
| NimbalystFREE | Kanban board for all open sessions; inline diff viewer; approve/reject from UI; mobile app | Free install at nimbalyst.com; connects via MCP |
| SecureContextFREE | Context Mode + credential isolation + SSRF protection + hybrid vector search | github.com/iampantherr/SecureContext |
MCP Servers
5 recommended| Server | Adds | Install |
|---|---|---|
| Context7FREE | Live, version-specific library docs injected before Claude writes code. Fixes hallucinated APIs. | Claude Desktop → Plugins → search "Context7" |
| GitHub MCP | Create repos, push code, manage branches, file PRs, do code reviews natively | claude mcp add --json github --scope user (external terminal; --scope user = all projects) |
| Playwright MCPFREE | Browser automation; Claude tests websites it creates; automated visual QA | Claude Desktop → Plugins → Playwright |
| Firecrawl | Web scraping that handles anti-bot measures; use in research and competitor workflows | API key + MCP config |
| Puppeteer MCPFREE | Screenshot feedback loops; Claude opens app, takes screenshots, iterates | Via npm |
GitHub Repos
6 worth bookmarking| Repo | Stars | Purpose |
|---|---|---|
| obra/superpowers | 177K | Forces planning; TDD; auto-promotes branches. In official marketplace. |
| Everything Claude Code | 140K | 48 agents, 182 skills, 68 commands; cross-session memory; works across Claude Code, Codex, Cursor |
| nexu-io/open-design | 19K | 72 brand design systems; Stripe, Airbnb, Tesla, Apple, Notion, Linear aesthetics |
| nateherkai/AIS-OS | , | WAT framework CLAUDE.md template + /onboard, /audit, /level-up starter kit. Free. |
| Karpathy Skills | 109 DL | One-file CLAUDE.md with 4 discipline principles. Add to every project. Free. |
| Roof Flow | , | 100 agents in swarms; federated comms across machines. Team-scale advanced use. |
Commands & Shortcuts
20+ found| Command | Does | When to use |
|---|---|---|
| ultra think | Allocates 32K token thinking budget for that one prompt | Architecture decisions, stuck debugging, complex strategy |
| /insights | HTML report on last 30 days of usage, graphs, patterns, anti-patterns | Monthly self-audit. Run it now. |
| /context | Token breakdown by source; lists every MCP tool loaded | Start of long sessions; when things feel sluggish |
| /cost | Actual spend for current session | Usage monitoring |
| /compact [keep X] | Compress conversation; optionally preserve specific decisions | At 60% context used |
| /clear | Full reset; CLAUDE.md and files persist, conversation doesn't | Between completely different tasks |
| /rewind | Roll back to a previous conversation point without starting over | When Claude goes the wrong direction |
| /schedule | Create cloud-hosted routine with plain-English description | Setting up automated recurring agents |
| /loop | Recurring prompt every N minutes in same session (up to 3 days) | Monitoring deployments; watching logs |
| /btw | Side question to Claude while it's mid-task; no interruption to main task | Non-blocking parallel queries |
| /init | Auto-generate CLAUDE.md by scanning project structure | Every new project setup |
| /model | Switch between Opus, Sonnet, Haiku | Cost vs. quality tradeoff |
| # (prefix) | Add instruction directly to CLAUDE.md mid-conversation | Capturing permanent rules without stopping work |
| effort high/low | Quality dial for that specific task | Complex tasks = high; quick edits = low |
| claude --worktree <name> | Isolated Git branch workspace for parallel session | Multiple features on same project simultaneously |
| Shift+Tab (twice) | Switch to Plan Mode | Before every new build |
| Shift+Tab (once) | Switch to Accept Edits mode (auto-approve file changes) | Active development without interruption prompts |
| Escape | Interrupt Claude mid-run immediately | When it's going the wrong direction; don't wait |
| Ctrl+Shift+V (VS Code) | Preview formatted markdown | Reading Claude's .md output in human-readable form |
Best Prompts & CLAUDE.md Patterns
10 verbatim1, Memory audit for skill discovery
"Based on all your memory, what skills would be relevant for me based on the work I do on a repeated basis?"
2, Planning question discipline
"Ask me clarifying questions until you're 95% confident before starting."
3, Self-improving system (goes in CLAUDE.md)
"When you hit an error, research the root cause, fix the tool, verify it works, and update the workflow SOP so this error never recurs."
4, Don't over-engineer (Karpathy principle)
"Build the simplest thing that could work. Make surgical changes only. Do not improve what was not asked."
5, Verification baked into to-do lists
"After each build step, take a screenshot and verify it looks correct before moving to the next step. Do not mark a task done until 95% confident."
6, Parallelization trigger
"You have [N] independent tasks. Spin up [N] sub-agents and run them simultaneously."
7, Business context persistence
"Save everything I tell you about my business into a business_profile.json file in this project so you never have to ask again."
8, Goal framing vs. step-by-step
"Build me a workflow that [describes outcome]. You figure out the how."
9, CLAUDE.md routing pattern
"If you need information about my business, read context/work.md. For my personal preferences, read context/me.md. For current projects, read context/projects.md."
10, Context window management reminder
"When context reaches 60%, compact the conversation keeping all architecture decisions and current task state."
Cost & Context Tips
10 tips- Disconnect MCP servers you aren't using for that session. Each loaded server costs ~18,000 tokens per message.
- Run sub-agents on Haiku for bulk reading, scraping, and data processing. Keep the main thread on Opus or Sonnet for synthesis only.
- Use direct API endpoints instead of MCP servers when you only need one or two functions. Saves the entire tool-definition token load.
- Batch multiple questions into one message instead of sending them separately. Each message reloads the full context.
- /compact at 60% context used. Use "compact but keep [specific decisions]" for precision rather than blind compact.
- CLAUDE.md should stay under 200 lines. Route to external sub-files for longer content. It loads as system prompt on every message.
- Context Mode plugin eliminates most context problems for long sessions (3-hour sessions vs. 30-minute crashes). Already installed.
- Claude Max plan ($100/mo) is the most cost-effective at heavy usage. Per-token overage costs exceed the plan delta quickly at high volume.
- YAML front matter in skill files = better triggering and token savings. Claude parses the metadata without reading the full skill body.
- /insights monthly audit reveals which patterns are burning the most tokens. Guaranteed to show something you didn't know.
Other Actionable Ideas
13 ideas- /ship-feature custom command: one slash command that commits, pushes, creates PR, merges, cleans up branch, and returns to main. Requires GitHub MCP.
- Sound notification hook: fire a sound when any Claude session ends. Enables running 15 parallel sessions without babysitting each one.
- Auto-commit hook: PostToolUse on Write/Edit fires git add -A && git commit -m "auto: claude edit". Every change checkpointed automatically.
- Session context briefing hook: UserPromptSubmit fires a script reading CLAUDE.md + git log + last session summary. Start every session with full context without re-reading files.
- PreToolUse file protection hook: blocks rm + .env + database files. Exit code 2 feeds the error back to Claude for self-correction. 100% deterministic vs. CLAUDE.md's ~98%.
- Chat transcript dump on Stop hook: full conversation window saved to file after every session. That file is your improvement data for better future prompt tuning.
- ChatGPT to Claude memory import: Settings → Memory → Import → paste Claude's export prompt into ChatGPT → paste results back. Imports your full ChatGPT preference history.
- business_profile.json per project: persistent memory for each business. Claude never asks for context it already has in that project.
- Modal for cloud deployments: webhook + cron triggers; only charges on execution; $35 free credit to start. Use after building workflows with Claude Code.
- @ file tagging in plan mode: type @filename to pull any file into the planning conversation. Especially useful for brand assets and reference documents.
- Obsidian Skills plugin: connects Claude Code directly to your Obsidian vault. Handles wiki links, embeds, callouts, and properties natively.
- Nimbalyst dashboard: free, open-source Kanban board for all open Claude Code sessions. Inline diff viewer. Mobile app to monitor sessions from your phone.
- em-dash PostToolUse hook: scans every file Claude writes for em-dashes and blocks the write with exit code 2. Makes the ban 100% deterministic rather than CLAUDE.md's ~98% compliance.
Videos Watched
10 videos1
From Zero to Your First Agentic AI Workflow in 26 Minutes
WAT framework; CLAUDE.md as onboarding doc; self-healing loop; business_profile.json pattern
2
Build a proactive agent workflow with Claude Code
Routines; /schedule command; GitHub event triggers; agent-on-agent review; live session steering from phone
3
How to Build $10,000 Agentic Workflows
WAT in depth; brand asset injection via @ tagging; newsletter workflow; value-based pricing framing
4
The Ultimate Claude Code Guide | MCP, Skills & More
/insights; /context; skills system deep dive; Nimbalyst; OpenClaw memory architecture
5
Hooks in Claude Code, Full Theory + Practical Use
Full hook anatomy; exit code protocol; /ship-feature command; blocking pattern with self-correction loop
6
Claude AI Full Tutorial: From Basics to Agentic AI
ChatGPT memory import; skills discovery prompt; Research Mode runs up to 45 minutes unattended
7
Claude Code: Build Your First AI Agent
CLAUDE.md 4-section structure; workflow files as plain English; effort levels; rollback; built-in web search
8
Nate Herk Channel Overview (multiple videos)
Context rot concept; CLAUDE.md routing pattern; agent teams setup; Context Mode + ClaudeMem combo
9
Claude Code Shorts + Pro Tips
47 pro tips; ultra think; /btw; /caveman; Remotion YouTube Shorts workflow; virality scoring
10
Discovery: 6 GitHub Repos That Change Everything
Superpowers; Everything Claude Code; Open Design; Karpathy Skills; Obsidian Skills; Roof Flow
Channels to Follow
Tier 1, Index immediately
@nateherk Nate Herk | AI Automation, highest signal-to-noise ratio; owns the WAT framework; 775K subs
@IndyDevDan IndyDevDan, deepest hooks and observability coverage; most technically rigorous
@TechWithTim Tech With Tim, 2M subs; broadest reach; excellent skills and MCP setup walkthroughs
Tier 2, Worth monitoring
Nuno Tavares Automated Marketer, best repo curation; surfaces the ecosystem layer
Greg Baugues 47 pro tips video is the best single reference card for power users
Sabrina Ramonov Surfaces lesser-known built-in commands most tutorials skip
Teacher's Tech Jamie, clearest beginner-to-agent progression; excellent CLAUDE.md structure walkthrough
Compiled 2026-05-31 · Private · ← Hub