Session 06

Claude Code·in depth

Program: Agentic engineering
With: Arndt Voges
From: Arndt Voges Consulting
Where: Remote

The tool itself, in depth

Claude Code in depthHands onClose

50 min Claude Code in depth, 25 min hands on, the rest together

0:00: Claude Code in depth
0:50: Break
0:55: In the session·write a command for something the team does often
1:20: Homework, close
1:30: End

Where we are

01Core concepts
02Preparing the codebase for agents
03Writing a good spec
04Feedback loops
05Workflow integration and code review
06TodayClaude Code in depth
07Orchestration and subagents
08Claude Code alternatives

The path of one ticket is complete. From here: the tool in depth, parallel work, the alternatives.

Five sessions of concepts.
Today: the tool itself.

What we cover today

Permission modesSandbox/btwStatus line

HooksSkillsSettings

CheckpointsDeny rules/rewindPlan mode

Code intelligence/contextCommands/goal

Agents, subagents, agent view and the rest: all part of the next session.

Part one

Settings andpermissions

Who decides what the agent may do.

Settings who decides what

Permissions: deny, ask, allow

.claude/settings.json

{
  "permissions": {
    "deny":  ["Read(**/.env)", "Read(./secrets/**)"],
    "ask":   ["Bash(git push *)"],
    "allow": ["Bash(pnpm test *)", "Bash(pnpm lint *)"]
  }
}

Deny is checked first, and blocks in every mode. The cheapest safety measure there is.

Six permission modes

Mode	Runs without asking	For
Manual	reads only	reviewing every action
acceptEdits	reads, file edits	iterating on code you watch
plan	reads, changes nothing	exploring before changing
auto	everything, a second model checks	long tasks, fewer prompts
dontAsk	only pre-approved tools	locked-down scripts and CI
bypassPermissions	everything	isolated containers only

Shift+Tab switches. Auto is the default on Pro, Max and Team plans, Manual on Enterprise. [VERIFY your plan]

Plan mode: read and think, change nothing.
The plan is the thing you review.

Ctrl+G opens the plan in your editor before you approve.

Part two

Skills, commandsand hooks

What the team writes once and uses every day.

Skill

Instructions and scripts the agent loads
only when a task needs them.

Almost no context until it is used, then the full instructions.

Commands are skills now

.claude/skills/commit/SKILL.md

---
name: commit
description: Commit with a summary of the changes
disable-model-invocation: true
---
Run pnpm lint and pnpm test first. Stop if either fails.
Summarize the staged changes in one line, then a short list.
Reference the ticket: $ARGUMENTS

/commit SHOP-1234 runs it. Files in .claude/commands/ still work the same way.

A central set of skills.

One repository, reviewed like code, distributed to every team as a plugin through a team marketplace.

Built-in skills worth knowing

`/code-review`

bugs in your diff

`/security-review`

vulnerabilities in your diff

`/verify`

checks against the running app

`/simplify`

cleans up the current diff

`/batch`

one change across many files

`/fewer-permission-prompts`

an allowlist from your history

Hook

A command that runs at a fixed point,
every time.

Format after an edit · tests before a commit · block a dangerous command. Rules that do not depend on the model remembering.

Before you build an MCP server,
check for a command-line tool.

gh, sentry-cli: the most context-efficient route. No CLI and no server? A small read-only server is about a day of work.

Part three

Working ina session

Undo, context, and daily tricks.

Resume when the context is still right.
Start clean when it is not.

/rename to name a session, /branch to try another direction, /resume to go back.

Checkpoints undo inside a session

Managing the context window

/context

See what fills it, as a grid.

/compact

Summarize on your terms: /compact focus on the API changes.

/clear

Empty context between unrelated tasks.

/rewind

Summarize from here compacts only part of the conversation.

Daily tricks

`@file`

point, do not describe

Paste a screenshot

images go straight in

`cat log | claude`

pipe data in

`!` command

run it yourself, output goes in

`/btw`

side question, never enters the context

Type while it works

messages queue, Esc interrupts

Ctrl+B

send a long task to the background

`/diff`

see what changed

Tracking tokens

/usage

Tokens and cost for the session, and what drives them.

Status line

Context and cost on screen, all the time.

Analytics

Usage per person and team, for the organization.

Even with a generous budget, where tokens go shows where the setup is weak.

CLI or editor extension

CLI

Full control and scripting
New features land here first
Runs anywhere, also in CI

Editor extension

Diffs where you already work
@-mentions and plan review
Lower barrier to start

The same agent. Pick per person.

Setup by level

Starting

The basics

CLAUDE.md, a test command, deny rules, /doctor to check it all.

Established

The team setup

Skills, hooks, a code intelligence plugin for Java or TypeScript, sandboxing.

Advanced

Beyond one session

Custom MCP servers, runs without a person at the keyboard.

In the session

Write a command the team will use.

20 minutes in small groups · for example a commit with a summary · each group shows one

Example homework, to be tailored to your team

Homework: three options

Individual

A plan before a change

A spec in plan mode for a ticket you own. What you changed in the plan, and whether the run followed it.

Team artifact

One hook and one skill

Committed, and shown working on a real ticket.

Cross-team

The central skill set

Each team contributes one skill; another team reviews it before it is merged.

Next session:
orchestration and subagents.

Arndt Voges Consulting · arndt@arndtvoges.com · arndtvoges.com

<stop>