AI Engineering in Five Minutes

The one-sentence version: AI output quality is set by the system around the model, not the prompt inside it. That system has five parts - autonomy, validation, telemetry, context, feedback - and seven moves. We call the whole thing the AI Engineering Operating Model.

1. Pick the discipline per task, not per person

AI is three workflows, not one, separated by a single variable: autonomy before review. AI-augmented - every line reviewed; for auth, payments, secrets, new patterns. Outcome-led - human defines testable outcomes, reviews at the PR; requires established local patterns. Product-led - human reviews feature behavior; low-risk code over ironclad automated checks only. The discipline follows the risk of the task, never the role of the person. When in doubt, default down. The Three Disciplines

2. Buy autonomy with validation

Autonomy is never free. Run it as an error budget: one nine = AI writes 90%, your 10% is a thick review layer. Two nines = your 1% must be spent precisely. 100% is a cliff - cycles compound at ~85-90% quality, and ten unreviewed cycles leave ~20%. Spend the human budget where judgment density is highest: known AI failure modes and architecturally dense code. Hand-written boilerplate signals an infrastructure gap, not a place to spend the budget. AI Code Nines

3. Burn and replenish, on telemetry

AI velocity spends quality like launches spend uptime. The telemetry is in your git log: absolute lines per day is the engine; the add/remove ratio is the direction. High ratio = feature burst. Low ratio at constant throughput = consolidation. Schedule the stabilization pass after every burst; turn its findings into written standards; feed them back into persistent context. Refactoring at AI scale doesn’t slow the engine - it points it inward. The Measurement Problem · The Stabilization Pass

4. Engineer what the model sees

The model mimics what is in front of it and invents what is missing. Control it three ways: an onboarding file of persistent rules - the LLM is a new hire, onboard it like one. Context models - capped, table-driven docs written for the model, tested on real tasks. Templates - complete verified exemplar systems, so generation becomes instantiation. The model doesn’t design your architecture; it instantiates one you’ve already verified. A Working AGENTS.md · Context Modeling · Template-Based AI Engineering

5. Make the change easy, then make the easy change

Two passes. First mechanical preparation - rename, extract, converge on the pattern, zero behavior change, cheap to review. Then the real change lands small in a clean zone. And manage the two resources underneath: plan the context window like server capacity (work the 20-60% band), and treat the session as disposable - the model cannot compartmentalize, the whole transcript votes, not your final decision - so end sessions at decisions and carry conclusions, never debates. Make the Change Easy · Context and Session Management

6. Review the session, not just the diff

The diff says what happened; the session says why. Three doors, in order: the ask (wrong or incomplete instruction?), the output (model drift despite a good ask?), the review (approved what should have been caught?). Findings become standards; standards become context; the next cycle starts cleaner. This loop is where the compounding lives. The Session Review

7. Scale vertically

The payoff. Horizontal scaling buys output with coordination tax; AI breaks that trade for the first time: one senior engineer’s judgment, flowing through the system above, produces multiple engineers’ output with zero new communication paths. The gates: a staff engineer at high nines keeps a ~3% human code budget that never reaches zero, and juniors still do the keyboard apprenticeship - review judgment is built by typing and owning mistakes, and you cannot review your way into judgment you never built. One org-chart consequence: held over years, the disciplines become working identities - the same IC title now spans opposite mixes of in-the-code and at-the-boundary work, and product splits into planners and builders - a dimension most ladders have no words for yet. Vertical Scaling · AI Impact on Career Ladders

The system, one paragraph

Choose the discipline by task risk. Set autonomy only as high as validation can pay for. Watch the add/remove ratio; replenish before the codebase forces it. Feed the model rules, context models, and templates so it instantiates your architecture instead of inventing one. Prepare mechanically, change cleanly. Replay sessions, bank the lessons into context. Spend the leverage on multiplying judgment, not headcount. Everything else on this shelf is the evidence.