romfast/clawd
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity
Files
72a62a409ffc92e7e8ea86457ded1431e8131d19
clawd/memory/kb/youtube/2026-03-03-claude-code-60-tips.md

10 KiB
Raw Blame History

Claude Code Expert Reveals 60 Tips Nobody Teaches

Sursă: https://youtu.be/AzmnaoVP8sk
Autor: Rahul Prihar (1,600+ ore experiență)
Durată: 41:30
Data: 2026-03-03
Tags: @work @growth #claude-code #productivity #ai-workflows

TL;DR

Rahul Prihar (1,600+ ore experiență) împărtășește 60 de tips avansate pentru Claude Code - de la workflow-uri cu worktrees și subagents, la headless mode pentru automatizări, skills custom, și tehnici de context management. Keynote: Claude Code poate fi folosit pentru orice, de la video editing la debugging DNA, limitat doar de imaginație.

Puncte cheie

🎯 Mindset & Capabilities

  • "Only limitation is your imagination" - a folosit Claude Code pentru 95% din editare video, automation, chiar dating apps (reverse engineered API, auto-reply, Telegram notifications)
  • Actualizează constant modelul mental despre ce poate face AI-ul - multe lucruri "imposibile" acum 2 luni merg perfect
  • $200/lună Claude subscription = $2,000 API usage → extrem de profitabil pentru background tasks
  • Notebook pentru limitări - notează ce NU merge, retestează cu fiecare model release

⚙️ Context Management

  • NU încarcă manual fișiere - lasă modelul să caute singur (evită poluarea context window)
  • Verbose mode ON (/config → verbose: true) - citește reasoning trace pentru insight-uri profunde
  • Separă planning de execution - plan într-o sesiune, execute în alta fresh (clear context)
  • Compactează doar când:
    • Model e focused pe aceeași task
    • Progresează clar, editează fișierele corecte
    • <3 compactions deja făcute
  • Fișiere scurte (sub 500 linii) > fișiere lungi - reduce read overhead dramatis
  • Status line (/statusline) - arată model, battery context, branch, unstaged changes

📁 CLAUDE.md Hierarchy

  • Stratificare: Root user → Project level → Subfolder specific → local.md (per-machine)
  • Progressive disclosure - doar relevant files loaded pe măsură
  • Reference files externe pentru long context (vocabulary, quirks)
  • Remove old rules cu fiecare model update - menține lean
  • WSL/macOS specific - local.md cu environment quirks

🤖 Subagents Strategy

  • Offload AGRESIV:
    • Online research
    • Codebase exploration
    • Log analysis
    • → Haiku subagents (cheap, fast)
  • Include "why" în subagent system prompt pentru filtered results
    • BAD: "How auth works"
    • GOOD: "How auth works for rate limiting because improving rate limiting"
  • Resumable subagents - poți relua un subagent să continue research
  • Fresh perspective - când stuck, cere subagent să exploreze

🛠️ Skills & Slash Commands

  • Custom skills pentru repeated prompts:
    • branch-commit-PR
    • log-explorer (macOS/Windows specific)
    • release workflows
    • localization-sync
    • seo-blog-writer
  • disable_model_invocation: true în frontmatter:
    • Skill manual-only, nu poluează context
    • Ideal pentru workflows you trigger yourself
  • Model per-skill (model: haiku) pentru well-defined workflows
  • Spec Developer skill - întreabă 20+ clarifying questions → plan de 500-700 linii covering edge cases

📊 Planning Best Practices

  • Start ORICE feature cu plan mode (Shift+Tab)
  • Spec Developer pentru features mari:
    • Massively improves initial plan
    • Reject initial plan → trigger Spec Developer → 500-700 line comprehensive plan
    • Peace of mind - înțelegi codebase-ul fără să citești cod
  • Spec Developer Explorer:
    • 3-5 Explore subagents APOI interview
    • Better questions când contextul e complex
  • Salvează toate planurile în tasks/ folder:
    • Undo features added on a whim
    • History analysis
    • Model comparison over time
  • Re-enter plan mode când features build on previous work (context carryover)

🔄 Parallel Workflows

  • 3-5 Claude Code sessions în paralel per project = "Factorio mode"
    • Always something to do
    • Spot bottlenecks across sessions
    • Improve pipeline, not just single-session speed
  • Worktrees (claude -w feature-1) DOAR pentru features overlapping în codebase:
    • Holds codebase static per feature
    • Evită race conditions
    • NU folosi pentru distinct parts (backend vs homepage)
  • Ghostty terminal cu hooks:
    • Notification când needs permission
    • Bell icon când done
    • Ctrl+Shift+R rename tab
    • Ctrl+D split window

🎨 Advanced Techniques

  • Brand new folder pentru fresh perspective când stuck în loop (avoid code bias)
  • Mermaid Diagram Generator skill pentru critical features understanding
  • Screenshots right size:
    • Nu prea mic (can't grep related text)
    • Nu prea mare (no idea what you're referring to)
    • Competitor references: big + zoomed detail
  • Speak agent's language:
    • Citește reasoning trace
    • Folosește terminology-ul lui Claude (ex: "CDN download URL" not "download URL")
    • Designers/developers have advantage - know the jargon
  • Voice dictation (HyperWhisper) - NU mai tastează, doar dictează prompts
  • Cross-project feature copy:
    • Add directory from another project
    • "Look through agents.codebase, identify Stripe integration, copy to HyperWhisper"

🔧 Tools & Integrations

  • Exa MCP server pentru recent packages fără training data
  • Official MCP servers (ex: trigger.dev) pentru heavy-use libraries
  • Claude-in-Chrome MCP pentru:
    • Forms filling automation
    • Anti-scraping sites
    • Taxes automation
    • Background repetitive tasks
    • Network debugging
  • Obsidian + Claude Code:
    • All notes = markdown on disk
    • Backlinks, graphic views, HTML generation
    • Obsidian CLI (recent release) pentru easier navigation

💡 Debugging

  • Când stuck in loop:
    1. Add logs throughout entire application
    2. Run, trigger issue
    3. Subagent găsește error în logs
    4. Fix
    5. Remove all logs
  • Document toate attempt-urile failed:
    • Start fresh chat cu document
    • Eventually găsește root cause
    • Smallest possible fix (Occam's razor)
  • Rewind tool (/rewind) > arguing cu modelul:
    • Restore code and conversation
    • Restore code only
    • Restore conversation only
    • Summarize from here
  • Isolate problem în clean environment pentru fresh perspective

🔒 Security & Sandboxing

  • HTTP proxy cu whitelist pentru sandbox:
    • Restrict Claude Code to approved domains
    • Prevent prompt injections
    • Company websites + trusted sources only
  • Dangerous skip permissions - NU folosi by default pe local machine

📈 Headless Mode

  • $200 subscription = $2,000 API usage → background processing la scară
  • JSON schema support pentru structured outputs
  • Use cases:
    • Receipt validation (hundreds of runs)
    • Hourly research automation
    • Course update tracking (changelog + transcripts → archiving/updating/new videos)
  • Cron + Skill + Headless pentru automated maintenance

🧪 Context Engineering

  • Audit context (/context) regular:
    • MCP servers: loaded on demand ✅
    • Memory files: delete old ones
    • Skills: some take 10-20% context! Use disable_model_invocation
  • Context anxiety (Cognition AI / Devin):
    • Model takes shortcuts near end of context window
    • Best: 1M token model, stay under 200k
    • Workaround: remind model it has plenty of context left
  • Long plans + compact:
    • Tick off completed items
    • Compact
    • Continue with tagged plan file
  • Start brand new chat când:
    • Context switching între taskuri
    • Model confused/editing wrong files
    • Already compacted 3+ times

Quote-uri notabile

"The only limitation to using Claude Code is your imagination. It basically feels like playing life on Minecraft creative mode."

"I find my biggest limitation is my bad assumptions of like, oh, the model would never be able to do that. But when I actually give the task to the model, it gets like 80, 90, 95% of the way there."

"Having one session, one task at a time is pretty slow. So what I usually do is I have 3, 4, or 5 instances per project running in parallel... There's literally always something that I can be doing."

"The bottleneck shifts from how fast can a single session complete a task to how well can I design my pipeline."

"I can explain how a lot of my codebase works at a pretty deep level, even though I've never read through the codebase, through the questions that Claude Code decided to ask me when developing plans."

"Context anxiety: the model begins taking shortcuts or leaving tasks incomplete when it believes it's near the end of its context window." — Cognition AI (Devin)

"Sweet. Awesome. Yeah. Second time, probably less excited, but still excited. 100 times every time you get the same victory, now it's called a job. That's fascinating. So, we need jackpots in our life to make us feel alive."

Acțiuni pentru Marius

  1. Implementează parallel sessions - 3-5 Claude Code sessions per project (Factorio mode)
  2. Ghostty terminal cu hooks pentru notifications când needs permission / done
  3. Spec Developer skill pentru features mari - peace of mind fără să citești tot codul
  4. Subagents AGRESIV pentru log analysis, codebase exploration (Haiku = cheap)
  5. Verbose mode ON - citește reasoning trace pentru înțelegere profundă
  6. Salvează toate planurile în tasks/ folder pentru history & undo features
  7. HyperWhisper dictation - stop typing, start dictating (included în Claude Code Masterclass)
  8. Mermaid diagrams pentru critical features understanding
  9. Headless mode + cron pentru automated research, receipt processing, course maintenance

Resurse

  • Claude Code Masterclass: https://clawd.io/masterclass (coupon: BIRTHDAY)
  • Claude Code Newsletter: Free, weekly tips
  • HyperWhisper: https://hyperwhisper.ai (free trial, Pro cu Masterclass)
  • Obsidian CLI: Recent release, easier agent navigation
  • Claude-in-Chrome MCP: Browser automation extension

Meta

Adăugat: 2026-03-03
Context: Link trimis de Marius pe Discord #echo