Kodokon kodokon.com

The developer in the AI era: staying critical and employable

Identify the skills that gain value as code generation becomes commonplace, and equip yourself to cultivate them deliberately.

8 min · 3 questions

Open this lesson in Kodokon

When a capability becomes commonplace, its market value drops and shifts toward what surrounds it. Written translation went through this, so did photography. For development, the commoditized capability is clear: producing correct code from a precise specification. What mechanically gains value are the two ends of the chain that AI doesn't cover. Upstream: specification - turning a vague, contradictory, political need into a well-posed problem; it's the skill every assistant user rediscovers the hard way, a prompt being nothing more than a specification whose every imprecision you pay for. Downstream: verification - review, testing, security audit, the ability to say "this plausible code is wrong"; it becomes critical precisely because the volume of plausible code explodes. And above both: architecture, the structural decisions whose effects are measured in years and which the model - with no memory of your organization and no responsibility for its advice - cannot carry.

The danger for a senior isn't replacement, it's undetected atrophy. Deep skills decline painlessly: nothing signals that your ability to debug without assistance or to design an API from scratch has eroded, until the day it's required. The answer is a deliberate, regular audit of what you actually exercise - as opposed to what you think you exercise. It's a delightfully contrarian use of AI: using the model to measure what the model makes you lose.

PROMPT
You are my technical career coach, direct and without flattery. Here is the honest log of my week of work with AI assistance:

- Tasks where I produced the work myself (AI as reviewer): [LIST]
- Tasks entirely delegated to the AI: [LIST]
- Code accepted without being able to explain it in detail: [LIST, even if it's uncomfortable]
- Technical decisions made this week: [LIST]

Analysis in three parts:
1. Skills actually exercised this week (the ones going up) versus delegated skills (the ones at risk of atrophy).
2. My comprehension debt: each item in the third list, with its concrete associated risk.
3. ONE 30-minute exercise for next week, targeting my most fragile skill, to be done without any assistance.

Format: assessment in 10 lines maximum. I prefer fair criticism to vague encouragement.
Weekly audit prompt - 15 minutes every Friday

Staying critical is also a skill to work on. The right stance is called calibrated trust: neither the systematic distrust that cancels the productivity gain, nor the blind trust that installs complacency. Calibrating means adjusting the intensity of verification to two variables: the cost of an error (a throwaway-script regex and an authorization policy don't deserve the same scrutiny) and the model's observed error rate on that type of task - which you'll only know by actually verifying, at first, everything it produces in your domain. Keep this count for a few weeks: you'll know where the model excels, where it hallucinates, and your vigilance will land exactly where it pays off.

PROMPT
From now on and for this entire conversation, adopt the role of mentor rather than executor. Permanent rules:

1. When I ask you for code, start by asking me to propose my approach; only present yours afterward, explicitly comparing it to mine.
2. When you correct me, state the general principle behind the correction, not just the one-off fix.
3. Once per response, ask me a question that checks I've understood, and wait for my answer before continuing.
4. If I simply ask you to "do it for me", remind me of this instruction and offer the guided version.

Confirm you've understood by restating these rules in a single sentence.
"Mentor mode" prompt - place it at the top of your work conversations

Employability, ultimately, boils down to a question every recruiter of the coming years will ask: what do you bring that the model doesn't? The defensible answer fits in three words seen throughout this path: specify, verify, decide - plus a method: keep learning deeply, with AI as a demanding tutor rather than a solution dispenser. A developer who understands what they deploy stays rare, and therefore expensive. That was the thesis of the first lesson; it's also the conclusion of the path.

Knowledge check

Make sure you remember the key points of this lesson.

  1. Which skill structurally gains value as code generation becomes commonplace?
    • Typing speed and knowledge of editor shortcuts
    • Specification and verification: posing the problem precisely, then judging what was produced
    • Exhaustive memorization of framework APIs
  2. What is comprehension debt?
    • The backlog accumulated on technology watch
    • The stock of code accepted without being able to explain it, whose cost is paid at the first incident
    • The classic technical debt due to design shortcuts
  3. What does calibrated trust in an AI assistant consist of?
    • Double-checking every suggestion, regardless of context
    • Adjusting the intensity of verification to the cost of an error and the model's observed error rate on that type of task
    • Granting trust after a month of use without incident