AI Toolchain & Pre-Flight Setup

Before you join the call: (1) Permissions pre-allowed — no 'Allow?' dialog during the interview. (2) Repo already warm — opened, indexed, AI has read the codebase. (3) CLAUDE.md or .cursorrules in place with project context. (4) Fresh context — /clear or new session, no stale conversation. (5) Parallel agents: OFF by default — serial is easier to narrate and review. Leave them off unless you routinely use them in daily work and can narrate them without confusion. (6) Drawing tool open in a separate window. (7) Phone timer visible — for the narration budget and T-minus checkpoints.

The tool should fit the task — the reasoning proves you chose intentionally, not habitually. 'I'm using Claude Code because it's fast for greenfield prototyping and I know the tool well.' 'I'm in Cursor because I need inline edits with full file context.' If your reasoning is strong but the tool is wrong for the task, the reasoning won't save you. Pick the right tool, then explain why.

Serial is safer and easier to narrate — default to it. If you routinely use parallel agents in daily work and can narrate them without losing context, keep them on. When you fire an agent, say what you're delegating and why: 'I'll have a sub-agent scaffold the API routes while I design the schema.' This signals you're orchestrating, not just reacting. Junior caveat: if you can't mentally track two code-generation streams, don't. 'I'm keeping this serial so I can review every line' is always defensible.

Call out MCP connectors explicitly: 'I'm connecting to the SQLite MCP server to set up the schema.' For libraries, explain the choice naturally: 'I'll use LangChain for the retrieval chain to save time on boilerplate.' Every tool choice should sound intentional, not automatic. Avoid real infrastructure (Postgres servers, cloud deployments) in a 60-minute prototype unless the task explicitly demands it — SQLite and local-first are faster and more reliable under interview pressure.

If the AI generates something you don't understand, say so and either read it aloud or rewrite it manually. 'Let me read through what it generated — I want to make sure the error handling is right.' Rule of Two: one fix-prompt, then write it manually. Blindly accepting AI output is a red flag. The interviewer is testing your judgment, not your prompting speed.

If your primary AI tool fails (rate limit, stale index, API 5xx): name it as external, try once more, and if it persists, switch to manual. 'Claude Code is rate-limited — I'm switching to manual in my editor, same principles apply.' The narration and reasoning habits carry over to bare-hands coding. The tool is an accelerator, not the engine.

Before the interview, ask your recruiter: 'Are AI coding assistants permitted in the live round?' Some companies ban them. Some require disclosure. Some interviewers view certain tools as 'cheating' while others encourage them. If the answer is no, adapt — the narration, architecture, and debugging principles all work without AI. If yes, confirm which tools are allowed. Never assume.