InterviewLab

AI agents

Real-world examples

Example: Build a new feature end to end

Prompt examples are illustrative; commands are verified against official docs.

Scenario: add a “forgot password” flow to an existing web app. The loop is the same everywhere — explore, plan, implement, test, review the diff, run a code review — but the mechanics differ per tool.

Workflow

  1. Explore the relevant code paths.
  2. Plan in the tool’s plan mode before any edit.
  3. Implement after approving the plan.
  4. Run tests and iterate.
  5. Review the diff with a built-in diff view or git diff.
  6. Code review the changes with the tool’s review feature.

Claude Code

1. Explore + 2. Plan. Start directly in plan mode (read-only; uses the built-in Plan subagent for research):

claude --permission-mode plan

Or toggle with Shift+Tab / /plan inside a session.

Explore how user accounts and email sending work in this repo, then plan a
"forgot password" flow: token generation, expiry, reset endpoint, and email.

Approve the presented plan to exit plan mode and start editing.

3–4. Implement and test. Claude edits files and runs commands with your approval (permission mode controls how much is automatic; auto is the default mode and auto-approves tool calls subject to background safety checks).

Implement the plan. After each file, run the related tests with
`npm test -- --grep auth` and fix any failures.

5. Review the diff:

/diff

6. Code review with the multi-agent reviewer (severity levels low to ultra; --fix applies fixes):

/code-review high

Codex

1. Explore in the TUI with the recommended low-friction combo for version-controlled folders:

codex --sandbox workspace-write --ask-for-approval on-request
Read the auth and email modules and summarize how a password-reset feature
would fit in.

2. Plan. Switch into plan mode (optionally with an inline prompt):

/plan Propose an implementation plan for a forgot-password flow: token
model, reset endpoint, email template, and tests.

3–4. Implement and test. Exit plan mode and let Codex work; workspace-write allows edits and commands inside the workspace (network is off by default).

Implement the plan step by step. Run `npm test` after each step and stop if
anything fails.

5. Review the diff (includes untracked files):

/diff

6. Code review — in the TUI against the working tree:

/review

Or non-interactively, with an explicit target:

codex review --uncommitted
codex review --base main

codex exec/codex review are the scriptable path (final message to stdout, progress to stderr), whereas the TUI flow is interactive.

Kimi Code

1. Explore in the TUI:

kimi
Read the auth and email modules and explain how a forgot-password feature
would integrate.

2. Plan. Start in Plan mode from the shell, or toggle in-session with /plan on or Shift-Tab. In Plan mode, Write/Edit are restricted to the plan file:

kimi --plan
Draft a plan for the forgot-password flow: token storage, reset endpoint,
email template, tests.

Exiting plan mode requires your confirmation — even in Ask When Needed (/yolo) mode, plan exits still ask.

3–4. Implement and test. After approving the plan, switch to Ask When Needed so routine edits and test runs don’t prompt one by one:

/yolo
Implement the approved plan incrementally. Run `npm test -- --grep auth`
after each change and fix failures before continuing.

5. Review the diff. There is no verified built-in /diff command; use shell mode (type ! in the empty input box) or ask the agent:

!git diff --stat && git diff

6. Code review. No dedicated review command is verified for Kimi Code — review via prompt:

Review the uncommitted changes as a senior engineer: correctness, security,
and test coverage. Report findings ordered by severity.

Last verified: 2026-09-07

AI coding tools evolve quickly. If you find outdated information, please open an issue or pull request.

View this doc on GitHub