claude in the terminal at least once
We do not continue until everyone has Claude Code working.
Checks connectivity, credentials and CLI version.
Shows the active model, accumulated cost and context used in the current session.
Sessions are per folder.
Make sure you are in
workshop-project before resuming.
Before, it took you 2 hours to write bad code.
Now it takes 2 seconds.
If you cannot explain what you want and you do not give it enough context, the agent will not know how to do it well either.
Simple tasks: rename, format, search.
Speed and cost over
depth.
80% of daily work.
The gap with Opus has shrunk a lot.
Architecture, plans, hard debugging, critical decisions.
Save it for when
Sonnet is not enough.
/cost with Bedrock · /usage with subscription — accumulated
spend for the session or subscription
github.com/rtk-ai/rtk — but it is a third-party tool and has telemetry enabled by default.
Run the project tests
mv ___CLAUDE.md CLAUDE.md
Run the project tests
Check if the project is ready
mv .claude/____rules .claude/rules
Check if the project is ready
The agent starts every session from zero.
It is like hiring someone new every
day.
The agent has persistent context: project conventions, how to validate changes,
what to avoid.
You never have to repeat it again.
Specific rules the agent always applies, organized by module or task type.
| Location | Scope |
|---|---|
~/.claude/CLAUDE.md |
All your projects — personal conventions |
<root>/CLAUDE.md |
The full project (versioned, shared with the team) |
<root>/CLAUDE.local.md |
The project, only your local copy (gitignored) |
<subdirectory>/CLAUDE.md |
Only that module — useful in monorepos |
.claude/rules/ |
Optional rules organized by file |
The closer to the code, the more specific.
Subdirectory overrides project,
project overrides global.
Keep the main CLAUDE.md short and import specific files depending on the context.
Local
overrides (@~/.claude/...) are loaded only on your machine — they are never
versioned.
Shows how much context you have used in the current session.
You can enable it
permanently in the status bar.
Claude runs it automatically when there is no space left.
It can lose critical context
that you have not saved.
You can run it manually.
When you need the agent to understand code from another module without opening a new
session.
Useful in monorepos or when a change crosses services.
If the other module has its own CLAUDE.md, Claude will read it
automatically when you add the directory.
Tools that Claude can use without asking.
Add them during the session
by answering "yes, always" or directly in settings.json.
Tools that Claude can never use, even if you ask.
Useful for
protecting destructive commands in production.
Run the project tests
Accept with "yes, always" to add it to the allow list
Run the project tests
| File | Scope |
|---|---|
~/.claude/settings.json |
Global — all your projects |
.claude/settings.json |
Project — committed with the repo |
.claude/settings.local.json |
Only you — not committed (add to .gitignore) |
Local overrides project, project overrides global.
Create a text file with a haiku in Japanese
Delete the haiku file
Undoes the last agent turn
In recent versions, it also undoes simple bash commands, but do not rely on
it...
Commands that run automatically in response to agent actions.
The
agent does not ask for them — they just happen.
| Event | Fires when... | Good for |
|---|---|---|
PostToolUse + Edit |
After every file edit | Linter / formatter — runs while the file is in focus |
Stop |
When the agent finishes its turn | Build / tests — waits until all changes are done |
Using the claude-code-guide agent, configure a hook in the settings.json of the
project to run the format tool after every file edit.
From now on the formatter runs on every edited file — without asking
mv ___settings.json .claude/settings.json
Every person on your team reinvents the same prompts.
Skills are the way to
share knowledge between sessions and between people.
A file with instructions that the agent loads when it needs it.
It lives in
.claude/skills/ and is activated with /name.
Claude can delegate subtasks to specialized sub-agents.
Each agent has its
own context, tools and CLAUDE.md.
| Prompts | Skills | Subagents | MCP | |
|---|---|---|---|---|
| What it adds | Step-by-step instructions | Reusable procedures between sessions | Task delegation with isolated context | Connectivity with external tools |
| Persistence | One conversation | Between conversations | Between sessions | Continuous connection |
| Contains | Natural language | Instructions + code + assets | Complete agent logic | Tool definitions |
| Loaded | Every turn | Dynamically, when relevant | When invoked | Always available |
| Includes code | No | Yes | Yes | Yes |
| Good for | Quick one-off requests | Specialized reusable knowledge | Specialized tasks with their own permissions | Access to external data and systems |
Do a code review of main.py
mv ___skills .claude/skills
Do a code review of main.py
| Skill | What it does |
|---|---|
/story-splitting |
Detects if a story is actually several stories in disguise |
/hamburger-method |
Cuts into layers, generates options, picks the one that gives feedback fastest |
/small-safe-steps |
Small increments, each one deployable on its own |
Turn it into a Skill.
That way the whole team shares it with no effort.
Generates skills automatically from a description.anthropics/skills → skill-creator
github.com/nikeyes/stepwise-dev
Plugin that groups research, plan, implement & validate
Complete workflows you can invoke with /name.
E.g.:
/research_codebase, /create_plan,
/iterate_plan, /implement_plan,
/validate_plan, /deep_research
Specialized agents that Claude delegates to automatically.
E.g.:
codebase-analyzer, web-search-researcher,
citation-analyst
Procedural knowledge loaded on demand.
E.g.: knowledge management in
thoughts/ or report generation with citations
If you want to join an existing working group, you can contact: