2026-03-27 14:51 Tags: https://youtu.be/TwkdDcO4vWQ?si=gQNbzBZa1qRwXQfQ
Conslusion:
- give it full permission (well~)
- good prompts contains:
- you specific goal
- give it example
- always ask open questions if you were an expert in xx what do you think
- always have plan mode
- don’t just accept the recommendation, if you don’t understand what they are doin, ask it, learn it! you need to know, you need to give it direction!
- use the skill! it can be better
- Context Window Management
- CLI is good
它甚至可以自己测试app，功能对不对。。。。 还可以直接告诉它commit and push to github… 推荐下载github cli and vercel cli 更新的话就告诉它commit and push to github.可以自动github and vercel都更新好。。。
In the terminal (main way)
When Claude Code is running, press Shift+Tab to cycle through permission modes:
- Auto-accept edits — Claude makes file changes without asking
- Plan mode — Claude describes what it would do but asks before every action
- Default — Claude asks permission for risky actions (file writes, shell commands)
3. Permissions System (Core Mechanism)
Claude Code operates with different permission levels, which determine how autonomously it can act.
Permission Levels
Default Mode
-
Claude asks for permission before every action
-
Includes file edits, command execution, etc.
-
Extremely slow and interruptive
Accept Edits Mode
-
Claude can edit files without asking
-
Still asks for permission before running system-level commands (e.g., installing tools)
This is the recommended starting point.
Bypass Permissions Mode
Enabled via:
claude --dangerously-skip-permissionsThis allows Claude to:
-
Modify files freely
-
Execute commands freely
-
Install tools, run scripts, etc.
Trade-off
-
More autonomy → faster workflow
-
Less control → higher risk
The transcript notes that most experienced users eventually move to bypass mode because:
Constant permission prompts destroy productivity.
4. The Most Important Skill: Prompting
The transcript emphasizes that prompting is not just important—it is the central skill.
Problem with Naive Prompting
Example:
“Build a kanban board”
This leads to:
-
Generic output
-
Missing features
-
Poor design decisions
Because:
- Claude fills in gaps arbitrarily
High-Quality Prompt Structure
1. Define the Goal (Not Just the Task)
Instead of:
Build a kanban board
You define:
-
Why the board exists
-
What it is supposed to enable
Example:
-
Organize past and future content
-
Track performance
-
Plan upcoming work
2. Provide Examples
Claude performs significantly better when given:
-
Screenshots (UI inspiration)
-
GitHub repositories
-
Concrete references
This reduces ambiguity.
3. Ask Open-Ended Expert Questions
This is one of the most important insights in the transcript.
You explicitly ask:
-
What am I missing?
-
What would an expert consider?
-
What are potential unintended consequences?
Why This Matters
AI enables you to operate in domains where you lack expertise. However:
You do not know what you do not know.
Therefore:
-
Claude must help fill those blind spots
-
But only if you explicitly ask
5. Plan Mode (Critical for Better Output)
Plan mode changes Claude’s behavior from:
“Execute immediately”
to:
“Clarify before executing”
Without Plan Mode
Claude:
-
Makes assumptions
-
Starts building immediately
-
Produces average results
With Plan Mode
Claude:
-
Asks clarifying questions
-
Forces you to define:
-
Features
-
structure
-
constraints
-
Key Insight
Plan mode is not just a feature.
It is a mechanism that forces thinking before execution, which is essential for non-technical users.
6. The Learning Trap (Very Important Section)
You can’t just accept, if you don’t understand , ask it!
7. Skills (How to Improve Claude’s Output)
Skills are described as:
Text prompts that modify how Claude performs tasks.
Two Types of Skills
1. Performance Skills
-
Improve quality of outputs
-
Example: frontend design skill
2. Workflow Skills
-
Combine multi-step processes into one command
-
Increase efficiency
Key Limitation of AI
Claude struggles with:
-
Taste
-
Design judgment
-
Creative decisions
Example: Frontend Design
Without skill:
-
Generic UI
-
“AI-looking” output
With skill:
-
Improved layout
-
Better aesthetics
-
More refined design
Important Clarification
Skills are:
-
Not automatic
-
Must be invoked explicitly or implicitly
Methods:
-
Slash command (guaranteed trigger)
-
Natural language (probabilistic trigger)
8. Context Window Management (Performance Control)
Claude has a limited context window (~1 million tokens).
What Are Tokens?
-
Roughly equivalent to words
-
Includes:
-
input
-
output
-
tool usage
-
skill usage
-
Problem: Context Rot
As context grows:
-
Performance degrades
-
Accuracy decreases
-
Outputs become worse
Observed Threshold
Around 20% of context capacity:
- noticeable performance drop begins
Solution: Reset Context
Command:
/clearThis:
-
wipes conversation memory
-
keeps access to project files
Why This Works
Claude can:
-
re-read files from the project
-
reconstruct context when needed
Best Practice
-
Keep sessions short
-
Reset frequently
-
Avoid unnecessary accumulation
Additional Tool
/contextAllows you to:
- monitor token usage
9. CLI Tools (Major Capability Expansion)
CLI = Command Line Interface tools that run in the terminal.
Why CLI Matters
Claude runs in the terminal.
Therefore:
-
it can directly control CLI tools
-
without heavy overhead
Benefits
-
Lower token usage
-
Faster execution
-
More reliable workflows
Example: Supabase
Instead of manually:
-
creating database
-
configuring auth
You can:
- ask Claude to do it via CLI
Example: Playwright
Playwright CLI enables:
-
browser automation
-
UI testing
Claude can:
-
open browser
-
simulate user actions
-
validate behavior
Key Insight
CLI transforms Claude from:
- assistant → operator
10. Automated Testing (Practical Application)
Using Playwright CLI:
Claude can:
-
generate test cases
-
execute them
-
simulate real user behavior
Example Actions
-
Create cards
-
Move elements
-
Verify UI behavior
-
Check interactions
Advanced Detail
“Headed browsers”:
-
visible browser windows
-
useful for debugging
vs
“Headless”:
-
runs in background
-
faster
11. Deployment Workflow (From Local to Public)
The transcript outlines a simple pipeline:
Step 1: GitHub
Commands:
-
commit → save locally
-
push → upload to GitHub
Claude can automate this.
Step 2: Vercel
-
Connect GitHub repository
-
Deploy project
-
Get public URL
System Flow
Claude Code → GitHub → Vercel → Live Website
Key Feature
-
Automatic redeployment
-
Every push triggers update
Error Handling
If deployment fails:
-
copy error
-
paste into Claude
-
Claude fixes it
12. The Four Core Levers (Final Synthesis)
The transcript ultimately identifies four ways to improve results:
1. Prompting
Defines:
-
what you want
-
why you want it
2. Skills
Improves:
- how Claude executes
3. Context Management
Controls:
- performance quality over time
4. CLI Tools
Expands:
- what Claude is capable of doing
13. Final Meta Insight
The most important conceptual takeaway:
Claude Code does not replace thinking.
It enables:
-
non-technical users to build
-
faster execution
-
broader exploration
But differentiation comes from:
-
problem definition
-
judgment
-
taste
-
decision-making
Critical Warning
If you:
-
always accept defaults
-
never understand tools
Then:
- you become indistinguishable from anyone else using Claude
Real Role of the User
You are not coding.
You are:
-
directing
-
deciding
-
refining
If you want, next step I can do something much more useful for your situation:
→ map this directly to how you should use Claude Code for your startup / job / portfolio building
That’s where this actually becomes powerful for you.