Claude Code
The Working Guide

01Install & first session

If you've got Pro/Max or an API key, you've got Claude Code. The install hasn't changed materially in months.

# install
npm install -g @anthropic-ai/claude-code

# verify
claude --version

# first session — drops you in the cwd
cd ~/Trading/NHous_TBOT
claude

Useful one-shot flags

02The daily-driver loop

The shape of a productive session — built around how you actually work: bursts, parallel projects, momentum-first.

03Slash commands

04Keyboard shortcuts (in-session)

05CLAUDE.md & memory hierarchy

CLAUDE.md is how Claude remembers who you are and what this project is, between sessions. Treat it like documentation that pays for itself every run.

The load order

# Most-general → most-specific (later overrides earlier)
~/.claude/CLAUDE.md          # global (every project)
<repo>/.claude/CLAUDE.md     # project hidden-dir
<repo>/CLAUDE.md             # project root
<repo>/path/to/sub/CLAUDE.md # nested wins inside its tree

What belongs in each layer

Project CLAUDE.md template for solo work

# <Project Name>

## What this is
One paragraph. What it does, who it's for, what shape the code is.

## Current state (update each session)
- Active focus: <the one thing>
- Blocked on: <or "nothing">
- Next session: <your future self will thank you>

## Key files
- `src/foo.py` — <one line about why this is load-bearing>
- `docs/ARCHITECTURE.md` — <the design doc>

## House rules
- <e.g.> Don't run migrations against prod. Ask first.
- <e.g.> `uv`, not pip. `pnpm`, not npm.
- <e.g.> Format on save. Test before commit.

06Models & cost routing

Three models. Different jobs. Route reads to Haiku, default to Sonnet, escalate to Opus when the problem really needs reasoning. Mid-session swapping via /model.

The decision tree

07Parallel sessions & worktrees

You run T-BOT, designdna, EntitleFlow, the portfolio. You can run two Claude Code sessions on the same repo at the same time — git worktrees make it safe.

Setup

# Terminal 1 — main branch session
cd ~/Trading/NHous_TBOT
claude
/rename tbot-main

# Terminal 2 — same repo, different branch, separate working dir
git worktree add ../NHous_TBOT-cci feat/cci-indicator
cd ../NHous_TBOT-cci
claude
/rename tbot-cci-feature

# When you merge the feature branch, clean up
git worktree remove ../NHous_TBOT-cci

How to use this without going insane

• Two sessions, max three. More than three becomes review-bandwidth limited. You can't keep up.
• One main thread, one long-running. Session A is the one you're actively driving. Session B is doing a 30-min refactor you'll review when it's done.
• Different verbs per session. "Investigate" vs "implement" vs "test". If both sessions are "implement", you'll context-thrash.
• Name them. Always. /rename within 30 seconds of opening or you'll lose track.

08Subagents

Subagents are scoped workers your main session spawns. They have their own tools, model, and prompt — your main context stays clean. This is where Claude Code stops being a chatbot and starts being a team.

Built-in subagents

Building a custom subagent

Drop a markdown file at .claude/agents/<name>.md (project) or ~/.claude/agents/<name>.md (global):

---
name: trade-log-analyst
description: Read-only analyzer for ICC trade journals and CSV exports
model: claude-haiku-4-5
tools:
  - Read
  - Glob
  - Grep
---

You are a focused trade-log analyst. Your job:

1. Read paper-trade journal entries and CSVs.
2. Summarize win rate, R-multiple distribution, drift from rules.
3. Flag cycles where SL was wider than 1.5x ATR.

Never modify files. Output Markdown only.

09Skills

A Skill is a reusable template that teaches Claude how to do a specific task your way. SKILL.md + supporting files in a folder.

Where they live

~/.claude/skills/<skill-name>/SKILL.md         # global
<repo>/.claude/skills/<skill-name>/SKILL.md    # project

SKILL.md frontmatter

---
name: nervahous-carousel
description: |
  Use when generating Instagram carousel copy for NervaHous.
  Triggers on "carousel", "IG post", "draft a carousel".
---

# NervaHous Carousel Skill

When asked to draft a carousel:
1. Pick a pillar (How-To Lab, The Stack, Concepts That Click, The Path, What Changed)
2. Hook in <7 words on the cover
3. 5-7 slides max, one idea per slide
4. CTA closer matching brand voice
5. Caption in NervaHous voice — warm, direct, curious

Always pass through the brand voice rubric in
~/NervaHous/Brand/voice-rubric.md before finalizing.

10Hooks

Hooks are shell commands Claude Code runs at lifecycle moments — before a tool call, after a file write, when a session starts. Use them to format code, block dangerous commands, log prompts, or notify yourself.

Where they live

~/.claude/settings.json              # global (every project)
<repo>/.claude/settings.json          # project
<repo>/.claude/settings.local.json    # local-only (gitignored)

Event types

Concrete hooks worth running

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "commands": [
          "if echo \"$COMMAND\" | grep -qE 'rm -rf /|:\\(\\)\\{|forkbomb'; then echo 'BLOCKED'; exit 1; fi"
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "commands": [
          "find . -name '*.py' -newer /tmp/last-edit -exec ruff format {} + 2>/dev/null; touch /tmp/last-edit"
        ]
      }
    ],
    "ContextFull": [
      {
        "matcher": "",
        "commands": [
          "osascript -e 'display notification \"Context full — run /compact\" with title \"Claude Code\"'"
        ]
      }
    ]
  }
}

11MCP servers

Model Context Protocol = how Claude Code talks to external systems. GitHub, Postgres, Slack, filesystem, browser. Lightweight servers Claude can call as tools.

Add / list / remove

# in-session
/mcp list
/mcp add github
/mcp remove github

# from CLI
claude mcp add github
claude mcp list

Scopes — which config wins

Servers worth wiring up

12settings.json reference

Minimal-but-real example

{
  "model": "claude-sonnet-4-6",
  "defaultModel": "claude-haiku-4-5",

  "permissions": {
    "allow": [
      "Bash(git:*)",
      "Bash(pnpm:*)",
      "Bash(uv:*)",
      "Bash(python:*)",
      "Read",
      "Write",
      "Edit",
      "Glob",
      "Grep"
    ],
    "deny": [
      "Bash(rm -rf:*)",
      "Bash(sudo:*)"
    ]
  },

  "env": {
    "NODE_ENV": "development",
    "PYTHONUNBUFFERED": "1"
  },

  "enabledMcpServers": ["github", "playwright"],

  "output": {
    "showCost": true,
    "showTokenCount": true
  },

  "context": {
    "autoCompactAt": 0.80
  }
}

13IDE integrations

14Cost & context discipline

Solo + ~$52K and counting = every dollar shows up. Cost discipline isn't about being cheap — it's about pointing the money at the work that pays back.

The /cost habit

• Run /cost after any task you'd describe as "expensive" — large refactor, big file scan, long Plan-mode walk-through.
• Once a week, run it at end of session for the rolling number. If it's surprising, ask "which task was the spike?"
• Set monthly soft budgets in your head per project. T-BOT $50/mo, OSS $30, portfolio $20. Recheck monthly.

Compact, don't churn

15Your specific workflows

162026 changelog highlights