Right now, I'm instructing an AI to write documentation for another AI. It works surprisingly well. In a few months, this entire workflow will be automated. The skill I'm building today has a shelf life shorter than the milk in my fridge.
The Graveyard of AI Skills
I've been working with AI professionally since GPT-3. In that time, I've watched entire skill sets go from cutting-edge to irrelevant. Here's the timeline:
Late 2024
Prompt Engineering
We spent months crafting elaborate prompt chains, few-shot examples, and system instructions. Then models got smart enough to internalize the patterns. The 20-page prompt template? Now a single sentence works better.
Early 2025
Manual Context Management
We built elaborate systems to chunk, summarize, and feed context to models with 8K token windows. Then context windows expanded to 200K, then 1M. All that infrastructure went obsolete overnight.
Mid 2025
RAG Pipeline Architecture
Retrieval-Augmented Generation was the gold standard. We built vector databases, embedding pipelines, and retrieval systems. Then models started handling longer context natively. The complex RAG stack became unnecessary for most use cases.
Late 2025
Agent Orchestration Frameworks
We designed multi-agent systems with specialized roles, handoff protocols, and coordination layers. Then single agents became capable enough to handle the entire workflow. The orchestration layer? Overhead with no benefit.
Why This Is Different from Normal Tech Churn
You might be thinking: "Tech has always evolved. React replaced Angular. Cloud replaced on-prem. What's new?"
When React replaced Angular, roughly 70% of your concepts transferred. Component thinking, state management, lifecycle patterns: all still relevant. The learning curve was incremental.
With AI skill decay, the problem you solved no longer exists. You didn't switch tools. The task evaporated. There's no transferable knowledge because there's nothing left to transfer it to.
This means getting really good at the current thing is almost a trap. The better you are at prompt engineering, the harder it hits when prompting becomes trivial. Expertise becomes an anchor.
You didn't switch tools. The task evaporated.
What Decays Fast
Tool-specific syntax and APIs
The exact way you call GPT-4 vs Claude vs Gemini changes every quarter. Memorizing API patterns is wasted effort.
Workarounds for model limitations
Every clever hack you build to compensate for a model's weakness gets obsoleted when the next version ships.
Model-dependent best practices
"Always use chain-of-thought" or "keep prompts under 500 tokens": these rules have a half-life of months.
Constraint-based architecture
Systems designed around current limitations (token limits, latency, cost) become over-engineered as constraints disappear.
What Compounds Over Time
Knowing what to ask
Problem decomposition (breaking a complex challenge into the right sub-questions) is a skill that only gets more valuable as models get more capable.
Recognizing wrong confident output
Models are getting better, but they still hallucinate with conviction. The ability to spot plausible-but-wrong output is a human superpower that compounds.
Understanding which tasks benefit from AI
Knowing when to use AI and when not to: this judgment improves with every project and transfers across every model generation.
Taste for "good enough"
Knowing when to ship vs when to iterate. AI makes it easy to over-polish. The skill of shipping at 80% and learning from real feedback is timeless.
Building feedback loops
Creating systems that learn from their own output (whether that's product analytics, user testing, or model evaluation) is the meta-skill that accelerates everything else.
Building a Learning System
If skills decay this fast, you need a system for learning, not just a list of things to learn. Here are five principles I follow:
The 80/20 Rule: Today's Work / Tomorrow's Tools
Spend 80% of your time on real work with current tools. Spend 20% experimenting with what's coming next. Not reading about it. Actually using it on a real problem.
Real Problems, Not Tutorials
Tutorials teach you the happy path. Real problems teach you the edges. Always learn by solving something you actually need solved.
Log What Broke
Keep a failure log. Every time an AI tool fails you, write down why. These patterns are worth more than any course because they reveal the actual boundaries of current capabilities.
Talk to People Who Work Differently
The best insights come from people in adjacent fields. A designer using AI differently than an engineer reveals patterns neither would see alone.
Build Disposable Things
Build prototypes you plan to throw away. The goal isn't the artifact. It's the understanding you gain from building it. Attachment to output slows learning.
The Uncomfortable Part
Here's something I don't often admit: even as the founder of an AI consultancy, I don't always feel like I'm keeping up. There are mornings when I open my feed and three tools I recommended last month have been superseded.
That feeling (the low-grade anxiety of perpetual inadequacy) is the new normal. And honestly? I think it's correct. If you feel like you're fully on top of AI, you're probably not paying close enough attention.
The difference between people who thrive and people who burn out isn't that the first group doesn't feel this. It's that they've made peace with it. They've accepted that mastery is temporary and learning is permanent.
What Actually Matters
Stop trying to be the expert. Start being the fastest learner in the room.
Velocity over expertise. The person who can pick up a new tool on Monday and ship something useful by Friday will consistently outperform the person who spent six months mastering last quarter's stack.
Intuition is the real skill. After enough cycles of learning, building, and watching skills decay, you develop a sense for what matters and what's noise. That intuition doesn't decay. It sharpens.
Stop trying to be the expert. Start being the fastest learner in the room.
How we work · Work · Insights