Teaching Junior Developers to Work With Agents, Not Against Them

A junior developer asked me last month whether using Claude Code would make them worse at coding. The honest answer is: it can, if they let it. The better answer is: used right, it is the fastest teacher they will ever have. Eight months of mentoring with agents has given me a workflow that accelerates juniors instead of replacing their thinking.

The Explain-Before-Accepting Rule

When a junior works with an agent, they cannot merge code they cannot explain. Not summarize, explain. Line by line if needed. If the agent produced code the junior does not understand, they either ask the agent to explain until they do, or they rewrite it themselves. This single rule prevents the skill atrophy everyone worries about.

The Debugging Drill

Once a week, I seed a bug into a file and hand it to a junior with strict instructions: do not ask the agent to fix it. Use the agent to explain what the code is doing, then find the bug yourself. The agent becomes a tutor, not a solver. Debugging is a muscle. Agents that do the debugging for you atrophy that muscle.

The Prompt-Writing Practice

Writing a good prompt forces clear thinking. I have juniors write the prompt for a task, review it with me, then run it. Half the value is in the conversation about why their first prompt was vague. A developer who can write a precise prompt can write a precise ticket, spec, or commit message.

A Pairing Template

15 min: Junior writes prompt for the task
5 min:  I review the prompt, suggest sharpening
20 min: Junior runs the agent, reviews every diff
10 min: Junior explains each change to me
10 min: We ship if understanding is solid, otherwise redo

One hour per task. Slower than letting them grind alone. Much faster than the six months of bad habits that come from letting juniors accept agent output uncritically.

The Long Game

Juniors trained this way become senior faster, not slower. They see more code, in more styles, than I did at their stage. The constraint is not exposure; it is comprehension. Slow the comprehension step and you get compound learning. The Teaching Tech Together guide has the pedagogy behind this pattern. I apply it to AI-assisted mentoring.

Agents do not replace mentorship. They multiply it.

The Skill the Industry Underestimates

Prompt precision is transferable. A developer who writes tight prompts also writes tight PR descriptions, tight bug reports, and tight standup updates. The underlying skill is the same: compressing intent into language someone else can execute on. I now evaluate junior candidates partly on the quality of their prompts, because that signal generalizes better than algorithm trivia. Language under constraint is a meta-skill; agents gave us an obvious way to practice it.