Browse the book & resources

Supporting material

Desk-researched interview 6: Skills gained, practice lost

Desk research, not original interviews. These question sets synthesize published accounts; no practitioner was interviewed for this book.

Question: Which skills are engineers gaining or losing? How are junior engineers getting feedback and practice?

Method: Desk research from published employee interviews, qualitative academic studies, practitioner accounts and operating guidance; no new junior interviews were conducted. Research snapshot: November 2026.

What the published accounts say

  • Anthropic surveyed 132 engineers/researchers and interviewed 53 of them in 2025. Employees described broader reach into unfamiliar codebases and faster prototyping, but some worried about losing direct debugging, exploration and the ability to judge generated code. A senior interviewee said junior colleagues asked them fewer questions, even though those colleagues got answers more quickly from Claude. These are mixed employee perceptions at an AI company, not measured skill assessments or proof of universal decline. Anthropic, How AI is transforming work at Anthropic.
  • Yu and Moon's 2026 study reports 14 semi-structured interviews in South Korea: eight students/recent graduates near workforce entry and six senior engineers. They identify a risk that routine junior tasks get absorbed into senior-plus-AI workflows, reducing opportunities to learn from mistakes. One senior's account differed: in that person's larger organization, juniors still had similar work. The sample is small and qualitative; it cannot establish the prevalence of lost skills or explain all changes in hiring. Crucially, its “juniors” are mostly pre-employment, not a sample of employed junior engineers. Yu and Moon, Who Will Become the Next Senior?.
  • Engineering lead Giordano Scalzo describes a junior's polished feature whose design and failure modes they struggled to explain. The published talk description proposes AI-off sessions, “Explain the Machine,” and “Break It” challenges to create feedback and debugging practice. This is a practitioner's illustrative account and proposed exercises, not a controlled evaluation of them. LeadDev, 29 September 2026.

Synthesis, not testimony

The gain is breadth, speed of exploration and AI-output evaluation practice; the possible loss is independent system understanding, debugging experience and human feedback. The evidence supports concern and design ideas, not a measured net skills balance. Primary outcomes are engaging growth and innovative learning; reliable judgment is a guardrail.

Protect junior practice without banning useful tools: pair on one unfamiliar incident, require a short explanation of design alternatives and failure modes, give feedback on the reasoning, then allow AI-assisted iteration. Periodically rehearse a small AI-off diagnostic task and compare progress over time. Track mentoring access and demonstrated skills, not just PR volume.

Questions to ask a practitioner

  1. Which concrete task can a junior now do independently that they could not do six months ago? Which can they no longer explain without assistance?
  2. How often do juniors get review comments about why a change works, and who provides them?
  3. What protected practice time, pairing or incident debrief builds the judgment needed to challenge AI output?

Follow-up desk-researched responses

1. What can a junior now do independently, and what can they no longer explain? No six-month before/after skill test is available for the same juniors. Anthropic employees describe reaching unfamiliar code and prototypes faster but worrying about lost independent debugging practice. In a distinct mixed-methods study, Feng, Yun and Wang observed 10 employed junior engineers working on an AI-assisted debugging task and found varying use patterns, including over-reliance and cautious avoidance in unfamiliar work. That is not proof that any one participant lost a previously held skill. Anthropic; Feng et al., 2026.

2. How often is “why” feedback given, and by whom? The Feng study had five senior engineers review anonymized junior code, prompt histories and reflections after the debugging task. They used those artifacts to infer reasoning and suggest mentorship through combined prompt-and-code reviews. This documents a research exercise, not the frequency or value of such feedback in normal PRs. Human feedback on reasoning is still needed; an AI answer alone does not establish understanding. Feng et al., 2026.

3. What protected practice actually exists? Google's SRE workbook describes new hires doing starter projects, guided labs, rollback drills and incident role-play before on-call, with mentors and escalation paths. That is a documented non-AI-specific training model; Scalzo's AI-off/explain/break exercises are a proposed adaptation. Neither source establishes that every junior engineer has protected time for these activities or that they improve measured skill retention. Sharing prompt histories for coaching should also be consensual and reviewed for sensitive content, not used as covert individual surveillance. Google SRE, On-Call; LeadDev, 2026.

Evidence gap: The cited sources do not measure junior skill growth longitudinally or establish that the proposed exercises improve production outcomes.