DIJJI Academy Engineering work in the AI era — what works, what breaks, what compounds.
The review bar was always a bundle of conditions enforced through the author's capacity to defend the choice. AI authorship pulled the bundle apart — some strands now pass for free, some can no longer be checked the old way, and new strands appeared that have no precedent. The bar is becoming several bars; each has to be picked up by name.
For two decades, seniors led greenfield (architectural judgment) and juniors handled brownfield (pattern mimicry). AI is strongest in brownfield and weakest in greenfield, inverting which kind of work needs which kind of engineer — and most teams are still staffing the old way.
The deployment cycle was tuned around an asymmetry — writing code was expensive, shipping it safely was the constraint. AI inverted that, and three of the cycle's most-trusted habits (small PRs, green CI, rollback as safety net) quietly stopped working the way they used to.
Engineering team memory used to live in three places: the code, the docs, and the people. AI thinned out two of them. The remaining one — documentation — got asked to do work it was never asked to do before, and most teams haven't yet rewritten what they write down to match.
Three assumptions about tech debt no longer hold under AI: debt was visible (it isn't), debt slowed implementation (it doesn't), and refactoring paid it down (it doesn't, by default). The bookkeeping system that worked for two decades stopped tracking the actual liability.
What it takes to develop as a junior engineer when AI tools are in the loop from day one. Pattern formation, code reading, mentor relationships, and calibration have all changed shape — and the conventional first-job curriculum stopped working.
AI tools made implementation cheap. Architecture didn't get cheaper. The brake that slow implementation used to put on bad design decisions is gone, the artifacts of engineering work moved upstream, and the bottleneck on a team's throughput is no longer where most teams are still budgeting.
Engineering metrics under AI: the link between the surface number and the signal it was tracking has decoupled. DORA reads healthier than the codebase, review metrics measure activity rather than depth, and the readings worth taking aren't on any dashboard yet.
Onboarding into a codebase with significant AI-authored history. Four habits that quietly stopped working — author interview, intent reading, style-as-thought, uniform trust — and what new joiners and mentors have to learn instead.
Incident response in a codebase with significant AI-authored history. Four practices — debugging, blast-radius assessment, runbook design, post-mortem analysis — quietly broke when the code stopped having authors who could be asked.
AI tools put the middle of the engineering org chart under more pressure than either end. The advice the field hands that layer — learn to prompt, upskill, move up the value stack — is mostly wrong or badly stated. What it actually takes.
Hiring engineers in the AI loop: which interview signals still mean what they used to, which got nullified, and which inverted. Plus the question every hiring manager asks about authorship — and the better one to ask instead.
AI-augmented engineering work lives or dies at three seams: prompt in, artifact out, review back. Each has a signature failure mode. The underrated skill isn't prompting, it's knowing when to throw the output away.
Standup, refinement, estimation, retro — four rituals that all changed shape when AI tools landed. Three of them quietly broke. The one that became the highest-leverage meeting on the board is the one most teams haven't reallocated for.
Three surfaces where AI tools change how software gets built: the sprint board, the skill mix, and the handoffs between humans and the tool.