Learning with AI tools

Juniors risk a quiet failure mode: the tool always writes the code, so you never build the mental models that make you an SDE2. Senior engineers multiply with AI; they do not disappear behind it. The job becomes specifying, verifying, and integrating — still engineering.

SDE2 expectation: faster delivery with equal or better quality, not faster delivery of mystery meat. In promotion packets, "used Copilot" is noise. "Cut migration time with agent + test guardrails while owning the cutover" is signal.

Interviewers increasingly ask how you use AI. A strong answer is a concrete workflow (plan → tests → agent → review), a failure story where you caught a bad suggestion, and how you protect learning. A weak answer is "I let it write everything." Pair this with coding interviews practiced without autocomplete so the skill remains yours.

Habits that protect growth

• Implement the first version of a new pattern yourself; use AI to vary it later.
• After an agent change, write the PR description from memory — if you cannot, you did not understand it.
• Use chat to quiz you ("ask me questions about this diff"), not only to produce diffs.
• Timebox "AI off" practice for fundamentals you still need in interviews and incidents.

Treat personal AI usage like a product with feedback:

• Time-to-green PR with vs without the tool on similar tasks (honest diary for two weeks)
• How often you revert agent commits entirely
• How often CI catches something you did not understand in review
• Whether your incident debugging skill is rising or atrophying

If revert rate is high, slow down: smaller tasks, more tests first, more plan mode. If you cannot debug without the chat pane open, schedule deliberate practice offline.

Teams can share a short "what worked" doc: good AGENTS.md snippets, forbidden prompts (prod data), and example PRs that used agents well.

Team norms

Agree explicitly:

• Secrets and proprietary data: what may leave the laptop / what tools are approved
• Whether AI-generated code needs disclosure in PRs (many teams want a short note)
• Required human review — AI review bots do not replace ownership
• License and compliance constraints for your industry

If policy is unclear, ask before pasting customer data or private keys into any cloud model.

Do not paste others' private code into tools that train on your inputs unless policy allows.

Common mistakes

• The tool always writes the first version of a new pattern. You never build the model.
• Debugging only with the chat pane open. Skill is atrophying.
• Pasting prod data "just this once." That is an incident.
• "Used Copilot" in a promo packet. Name the cutover you owned.

Check yourself

1. You cannot write the PR description from memory. What does that mean? You did not understand the diff. Do not merge until you can.

2. Policy is unclear and the fixture has customer emails. What do you do? Ask before you paste. Scrub PII. Use the approved enterprise tier.

Key takeaways

• Protect skill growth: understand before you merge; practice without the tool on purpose.
• Eval your own workflow: revert rate, CI surprises, incident skill.
• Follow team policy on secrets, disclosure, and approved vendors.
• Promo signal is the cutover you owned, not the tool you used.

Further learning

Anthropic on Claude Code best practices — instruction files, planning, verification.

Claude Code: Best practices for agentic codingTips and patterns for getting the most out of Claude Code, from configuring your environment to scaling across parallel sessions.Claude Code Docs Simon Willison's WeblogSimon Willison’s WeblogSimon Willison’s Weblog