Claude Code: Build Safely
Use Claude Code as a capable collaborator, not an autopilot. Learn how to give context, constrain scope, review proposed work, verify every change and protect secrets or important files.
Claude Code works inside a real folder.
What if a coding assistant could read the whole project—not just one pasted snippet? That is powerful precisely because the folder, permissions and final decisions still belong to you.
Claude Code is an AI-powered coding agent that runs in a terminal. It is not a programming language, framework or replacement for Git. It can inspect a codebase, explain files, propose edits and run commands you allow. Developers use it for careful debugging, repetitive changes, test writing and codebase research.
It needs internet access, a supported installation and currently eligible Anthropic access; use the free installation guide. Because it works inside real folders, you must choose the folder carefully, understand permissions and review what changed.
Before the first prompt
- Follow the current official installation and authentication guide for your platform.
- Open only the project folder you intend to work on—not your entire home directory.
- Make a Git commit or recoverable backup.
- Remove secrets from prompts and tracked files; use environment variables and ignored local files.
- Know the command that proves the project still works.
Outcome: Add a filter to the resource list.
Scope:
- Work only in src/resources/ and its tests.
- Keep the existing visual design and data format.
Before editing:
1. Inspect the relevant files.
2. Explain the current data flow.
3. Propose a short plan and wait for approval.
Done means:
- Keyboard and pointer users can filter.
- Empty results are explained.
- Existing and new tests pass.
Safety check
Use one repeatable loop.
- Inspect: ask what files, behaviour and tests currently exist.
- Plan: agree on a small change and its acceptance checks.
- Change: let Claude implement only the approved scope.
- Verify: run checks, inspect the diff and exercise the real user flow.
- Record: commit a coherent change with the reason, not just the file list.
Inspect the form submission flow without editing anything.
Report:
- the entry point and data path,
- existing validation and tests,
- likely cause of the duplicate submission,
- the smallest safe fix,
- exact commands and browser steps that would verify it.
Stop after the plan.
What belongs in CLAUDE.md
- Short project purpose and architecture.
- Setup, formatting, test and build commands.
- Important conventions and directories.
- Actions that require approval.
- Definitions of done that apply repeatedly.
Do not put credentials, temporary task details or an enormous copy of general documentation there. Keep instructions specific enough to change decisions.
Why ask for inspection before a plan?
A confident plan based on guessed architecture is still a bad plan. Grounding the proposal in actual files, tests and behaviour reduces unnecessary edits and makes the review concrete.
“Done” needs proof.
Read the diff. Run syntax, unit and build checks. For a UI change, open the real page at desktop and mobile sizes, exercise the interaction and check the console. A green build cannot prove that the button is readable or that the workflow makes sense.
Never delegate blindly
- Destructive file or database operations.
- Publishing, payments or messages to real people.
- Security, privacy or legal decisions.
- Handling secrets and production credentials.
- Claims that require current external evidence.
Map before you modify
Inspect a small project and explain its data path before proposing work.
Make project instructions useful
Write a focused CLAUDE.md with commands, conventions and boundaries.
Reproduce, fix and test
Prove a bug, make the smallest repair and verify edge cases.
Human-controlled Git review
Plan a change, inspect its diff and keep commit decisions human.
Use extensions safely
Review a read-only subagent and an MCP tool with a small blast radius.
Ready to use it on the capstone?
The complete note goes deeper.
The full authoring pack is complete locally and is being prepared for protected delivery. It will include the remaining explanations, practice checks, all staged projects and the final combined build. Payment is not active yet.
- Current setup and authentication
- Project context and CLAUDE.md
- Inspect, plan, change, verify
- Permissions and security
- Tests and debugging
- Git checkpoints
- Agents and MCP boundaries
Primary references reviewed 13 August 2026: Claude Code overview · Claude Code setup · Claude Code authentication · Claude Code permissions · Claude Code best practices. Explanations and examples are original KODE Ń VIBE teaching material.