https://www.youtube.com/watch?v=Y09u_S3w2c8
00:00 - Every Level of Claude Code Explained in 39 Minutes 00:41 - Level 1: The One Thing You Need To Know 03:25 - Level 2: Personalizing CLAUDE for better responses 10:25 - Level 3: Slash Commands, Skills & Hooks - Repeatability 21:21 - Level 4: Connecting Claude Code to Your Apps (MCPs) 26:18 - Level 5: Move from Executor to Supervisor 29:34 - Level 6: Agent Teams 35:57 - Level 7: Fully Autonomous Systems
Level 1 : Prompt
- first run in plan mode
- then commit
The askUserQuestion is used to build context. By asking questions on assumptions.
The AskUserQuestion tool is an interactive feature within Claude Code (introduced in v0.21) that allows the AI to pause execution and ask the user clarifying questions when it is uncertain about the user's intent or needs further direction. It is designed to gather user preferences, define requirements, or make implementation choices, reducing hallucinations and improving the accuracy of code generation.
Key Features & Functionality:
- Interactive Interface: Presents between one and four questions at once, supporting single-select, multi-select, and custom input ("Other").
- Plan Mode Integration: Active during "Plan Mode," where Claude asks questions to confirm requirements before executing changes.
- Clarification Capabilities: It can display markdown snippets, including code or diagrams, to help the user understand the context of the question.
- Error Handling: Users can respond with custom text to provide more details if the predefined options are insufficient.
Usage in Workflow:
- Preventing Assumptions: It forces a shift from AI assumption-making to a confirmation-seeking behavior.
- Best Practice: Developers often prompt Claude to "ask me if anything is unclear" to trigger this tool, especially before starting large tasks.
- Limitations: The tool works primarily at the command level (main thread) and cannot be used within subagents.
Common Use Cases:
- Asking for verification on which files to modify.
- Probing for environment details (browser, OS, network).
- Reviewing work flows and seeking approval before writing final code
Level 2 : Claude.md (follows your personalization & rules )
- understand rules
- how you want to work with it ( tech stack , preferences , mistakes you do not want to repeat )
Golden Rule : The golden rule: Short. Specific. Only what Claude can't figure out itself.
is ready by /init command
Level 3 : Repeatability (Slash Commands, Skills & Hooks)
- Slash commands
- is saved prompts which you can reuse
- it is like pressing a button
- Skills
- https://skillsmp.com/
- Background knowledge Claude loads automatically when it's relevant (based on project description and skill description) . You don't trigger these — Claude just knows to use them when it needs to.
- still represented as slash commands
- Unlike commands, skills can include a whole folder of supporting files — example posts, style guides, reference docs — giving Claude much richer context to work from
- brand-voice : your tone, banned words, sentence style, plus a folder of 10 example posts that nail your voice. Claude pulls this in whenever it's writing any content, without you asking
- You never type /brand-voice. Claude just knows "I'm writing content, I should check the brand voice skill." And because skills are folders not just single files, you can pack in all the context Claude needs — examples, templates, reference material — not just a list of rules.
- Use skills when you want to sometimes provide instructions to your primary agent with a relevant skillset. Use slash commands when there are things you specifically know you’ll want to invoke at certain points.
- Hooks - these don't require a brain (llm tokens)
Hooks are automatic triggers / mechanical checks that fire when Claude does something. It’s for stuff a bash script can do without needing Claude's brain.
inside .claude create a settings.json.
Level 4 : MCP Servers
https://github.com/wong2/awesome-mcp-servers
add a .mcp.json
or you can remove above and run below command.
Level 5 : Human move into supervisor role with GSD
GSD = Get Shit Done
Removing ourselves from the thinking overhead - Project Frameworks - GSD
Install it with one command:
npx get-shit-done-cc
It leverages the same askUserQuestions feature we see used heavily in the planning mode - however the key difference here is the level of detail of the breakdown of the plan -> which helps solve the biggest problem with LONG claude sessions - context rot.
Run through each project phase using the plan, execute and verify commands in sequence
- /gsd:plan phase X
- /gsd:execute phase Z
- /gsd:verify phase Y
Context is pulled for each phase from the overall project documents
- ROADMAP.md
- REQUIREMENTS.md
- STATE.md
as well as phase specific documents.
Install it with one command:
npx get-shit-done-cc
It leverages the same askUserQuestions feature we see used heavily in the planning mode - however the key difference here is the level of detail of the breakdown of the plan -> which helps solve the biggest problem with LONG claude sessions - context rot.
Run through each project phase using the plan, execute and verify commands in sequence
- /gsd:plan phase X
- /gsd:execute phase Z
- /gsd:verify phase Y
Context is pulled for each phase from the overall project documents
- ROADMAP.md
- REQUIREMENTS.md
- STATE.md
as well as phase specific documents.
Context Rot
Framework like GST helps manage context rot. it keeps context broken into smaller files.
Level 6 : A Team of agent
Instead of 1 claude agent doing everything (researching , writing etc ) we have a team of agents , split the job into separate sub-agents.
Running multiple agents - increase your leverage
Summarised very well in this Reddit thread
It is all about context rot and context isolation.Level 7 : Fully Autonomous Pipelines : Ralph Loop
No comments:
Post a Comment