roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Build software with an AI coding agent and keep the result correct, secure and yours to maintain. Covers the agent workflow, reviewing generated code, prompting and context, and Git as the safety net.
491 of 491 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 491 new. 0 cut, outside your plan.
Interviewers usually test how you work with coding agents inside rounds built for something else: in the behavioral conversation about something you shipped with an assistant, in a take-home that allowed AI tools and is then read line by line, and in coding rounds that let you use an assistant on screen, where the interviewer watches what you hand over and what you check before accepting it.
The probes are about failure. The agent reports the suite green after rewriting the assertion that failed. A generated test passes because it was derived from the code it checks, so it confirms the bug too. A suggested dependency exists only because someone registered a name assistants tend to invent. Late in a long session, an instruction given at the start no longer holds. An agent ran git reset --hard, and the question is what can be recovered: commits through the reflog, staged content as dangling objects until garbage collection, and edits that were never staged not at all.
What separates levels is where the safeguard lives. Early on it lives in habits: small tasks, every diff read, a commit before each run. From senior level up it is expected to live in the setup: a team-maintained instruction file, permissions that keep the agent away from secrets and destructive commands, a hook that runs the tests before the agent may call the work finished, and a view of what a short tool trial can and cannot prove.
The plan opens on practice rather than product. Vibe coding as a working mode, cutting work into pieces an agent can finish and you can check, comes first with the assistant inside the edit-test-review loop, because both hold for every tool that follows.
AI Coding Tools is the widest stretch, and it runs from the terminal toward the browser. Claude Code leads and goes deepest: its agent loop and the permissions around it, memory files, and the hooks the harness runs whatever the model decides. Gemini CLI and Codex are agents of the same kind. Then the editors take over, where Cursor's rules and Copilot's instruction files are the same standing-context idea in another form, and Copilot's agent mode is the agent loop inside the IDE. v0 and Lovable close it, building whole apps from a chat with the code behind a preview, where the review habit from the first module is hardest to keep.
Prompting & Context then explains what the tools do. Prompt engineering matters here mostly for evaluation and iteration: telling a prompt that helped from one that happened to work once. Context engineering is the mechanism behind the memory files and session advice: what fits in the window, what compaction keeps and drops when a session is summarized, and why an instruction's position in the window changes how much weight it carries.
The plan ends on Git, the net under everything above it. Staging and commits come first, down to partial staging, which lets you keep the hunks of an agent's diff you checked and drop the rest. Branching and merging gives each agent task a place to go wrong in isolation, and conflicts once parallel agents touch the same files. Recovery closes the plan: reset, restore and the reflog, the commands that decide whether a bad agent run costs a minute or a day.
Where a loop lets you code with an assistant on screen and the date is close, rehearse in the tool you will use, saying aloud what you check and why you accepted each change, because that narration is what gets graded. Cutting importance back to must know helps most here, because the tool modules carry a long tail of product detail, and the modules for tools you will not use can stay unopened. If you learned to build software through these tools rather than before them, Git is the gap, printed last but worth reading first. Recovery commands are safe only once you can say which of the three trees a change lives in, and a study order of by topics keeps staging ahead of reset. Senior engineers who run an agent daily know most of the tool modules already; their questions sit in the first module, context engineering and Claude Code's [security and governance](/topics/ai-claude-code-security) material, where a personal workflow has to become a team rule. The Git here stays on your machine: remotes, pull requests and rebasing, where an agent's work meets the team, belong to the [Git & GitHub roadmap](/roadmaps/git-github). And the plan is about using agents someone else built; writing your own, with a tool-calling loop and an evaluation harness in your own code, belongs to the [AI Agents roadmap](/roadmaps/ai-agents).
Where a loop lets you code with an assistant on screen and the date is close, rehearse in the tool you will use, saying aloud what you check and why you accepted each change, because that narration is what gets graded. Cutting importance back to must know helps most here, because the tool modules carry a long tail of product detail, and the modules for tools you will not use can stay unopened.
If you learned to build software through these tools rather than before them, Git is the gap, printed last but worth reading first. Recovery commands are safe only once you can say which of the three trees a change lives in, and a study order of by topics keeps staging ahead of reset. Senior engineers who run an agent daily know most of the tool modules already; their questions sit in the first module, context engineering and Claude Code's security and governance material, where a personal workflow has to become a team rule.
The Git here stays on your machine: remotes, pull requests and rebasing, where an agent's work meets the team, belong to the Git & GitHub roadmap. And the plan is about using agents someone else built; writing your own, with a tool-calling loop and an evaluation harness in your own code, belongs to the AI Agents roadmap.