Request one Portfolio ↗
Life with a product 9 chapters about 5 min 17 Oct 2026 From request #1

Your first week with Claude Code

Join a small team on Monday and ship on Friday, pairing with Claude Code: what it does when you ask, what it asks you back, and the habits that make someone good with it.

Start full screen ↗ Play it here ↓

Drag to look around · arrow keys step through · T for the text view

A week lived from the developer's chair. The codebase is a small city, Claude Code a light that reads its streets, asks at the gate and walks a dotted plan; a tech lead and a support colleague speak in their own voices, and every day has a choice that the scene answers.

The brief it was built from

Learn Claude Code by living a first week with it, from a first prompt to expert habits: how it reads a codebase, its memory file, checking its work with tests, permissions and modes, planning, managing context, extending it, and shipping. Fully narrated, so it can also be followed by ear.

What is real, and what is a scenario

Real data
  • What Claude Code does, as stated in Anthropic's documentation: it reads your project's files as needed, edits files and runs commands, and works with git and the gh command line tool to write commit messages and open pull requests
  • CLAUDE.md is read at the start of every session, and /init writes a starter file
  • The permission prompt (yes, yes and don't ask again for a command, no), the permission modes (manual, accept edits, plan, auto, cycled with Shift+Tab), and that newer versions start interactive sessions in auto mode, where a classifier reviews actions
  • Plan mode: Claude reads, explores and writes a plan, and does not edit your code until you approve it
  • The context window fills as a session goes on and performance degrades as it fills; /clear starts fresh, /compact summarises, and Claude Code compacts automatically as it approaches the limit
  • Escape stops Claude; Escape twice on an empty prompt opens rewind, which restores Claude's file edits and the conversation but not changes made by terminal commands
  • Hooks (for example a formatter after every edit, on the PostToolUse event), subagents in their own context window, MCP servers that connect tools such as an issue tracker, and claude -p for scripts and continuous integration
  • The habits taught follow Anthropic's best-practices page: explore, then plan, then code; give Claude a way to verify its work, such as a failing test that reproduces the issue; be specific; course-correct early
  • India Standard Time is five and a half hours ahead of UTC, so an order between midnight and 5:30 am IST falls on the previous UTC date
Scenario & assumptions
  • Dabba, its four-person team, its 42-file codebase and its line counts are invented for the story; the city is a picture of that invented codebase, not of a real one
  • Arun, Divya and the customer are roles in a labelled scenario; every line they say is written for the story. Claude Code's replies are shown in a terminal in real life; here its lines are voiced so the run can be followed by ear, and their wording is illustrative
  • The terminal on the desk is an illustrative, simplified picture of a Claude Code session, not a screenshot: the layout, the mode labels and the tool names are approximations
  • The scenario keeps you in manual mode for the week so the permission prompts can be shown; newer installs start in auto mode, and your organisation's settings may differ
  • Every time on the clock, the hours for Wednesday's feature (plan and build, and the Thursday rework in the fast style), the test counts and the pull request number are a scenario, not a benchmark
  • The confidence and context meters are illustrative: the context meter shows the idea of a fixed-size working memory filling and being compacted, not a measured percentage
  • The 'calls made the way an experienced user would' count in the outro reflects the choices this page offers, and is not an assessment of skill
  • Five minutes is the autoplay pace; making the choices takes longer

What you leave with

  • Ask Claude Code about the code before reading it yourself, and say what you want and where to look
  • Keep CLAUDE.md short, true and stable; it is guidance, and hooks are how you enforce a rule
  • Have it write a failing test before it fixes a bug, so its work can be checked
  • Read every command it asks to run; allow only safe, repeatable ones for good
  • Plan anything that touches money or many files, and question the plan before approving it
  • Clear the context between unrelated tasks, rewind wrong turns, and read every line before it ships

Sources

  • Claude Code overview What Claude Code is: reads your codebase, edits files, runs commands; works directly with git (commit messages, branches, pull requests).
  • Claude Code quickstart Start with claude in your project folder; Claude Code reads your project files as needed, so you do not have to add context by hand.
  • Set up Claude Code Installation, and the account types that include Claude Code (the free plan does not).
  • Permissions The permission prompt and its options; reads and read-only commands need no approval in manual mode; 'don't ask again' for a command is saved per project.
  • Permission modes Manual, accept edits, plan, auto and the others; Shift+Tab cycles; newer versions start interactive sessions in auto mode; plan mode and its approval step.
  • Memory: CLAUDE.md files CLAUDE.md is loaded at the start of every session; /init generates a starter file; it is context, not enforced configuration.
  • The context window Automatic compaction as the limit approaches; /clear and /compact.
  • Commands /init, /clear, /compact, /context, /memory, /rewind.
  • Checkpointing Rewind restores Claude's file edits and the conversation; changes made by terminal commands are not tracked; not a replacement for version control.
  • How Claude Code works: interrupt and steer Escape stops Claude immediately; Escape twice on an empty prompt opens rewind.
  • Subagents Specialised assistants, each with its own context window, tools and permissions.
  • Hooks guide User-defined shell commands at lifecycle events; the formatter-after-edit example on PostToolUse.
  • Model Context Protocol in Claude Code Connecting tools beyond the built-in set, such as an issue tracker or a database.
  • Skills Reusable prompts as skills; custom commands have been merged into skills.
  • Headless mode claude -p runs one prompt non-interactively, for scripts and continuous integration.
  • Best practices for Claude Code Explore first, then plan, then code; give Claude a way to verify its work (a failing test that reproduces the issue, then the fix); provide specific context; course-correct early and often; manage context aggressively. The older engineering-blog address redirects here.
  • Create pull requests (common workflows) Claude Code opening pull requests, using the gh command line tool.
  • Time zone: India Standard Time (UTC+05:30) The offset behind Tuesday's bug.

3D with three.js (MIT) · Narration voice: Kokoro-82M v1.0, voice af_heart (kokoro-onnx 0.4.7). Apache-2.0 (Kokoro-82M weights).

Keep going

More experiences

All experiences →