If your company spent the last two years building a prompt library, a lot of it is already dead weight. Not because your team did anything wrong. Because the models got better.
That sentence sounds backwards. Let me show you exactly what I mean.
The Upgrade That Broke the Playbook
The newest generation of frontier models, Claude Fable 5 and Opus 5, verify their own work automatically. You give them a task, they do it, they check it, they correct themselves. That behavior is built in now.
Two years ago it wasn't. So every careful AI operator, myself included, wrote prompts that compensated: "Work step by step. After each step, verify your output. Check your work before continuing."
Run those same prompts on the new models and something strange happens. The model verifies its work because you told it to, then verifies again because that is what it does by default. Your instructions push it into long, redundant checking loops. Tasks that should take one pass take five. You pay for every one of those passes, in time and in tokens.
Read that again, because it is the whole point: the prompt failed because the model improved. The instruction was a crutch for a weakness that no longer exists, and now the crutch is tripping the runner.
Prompts Are Workarounds, and Workarounds Expire
Here is the pattern underneath the example, and it is the part most companies miss.
A prompt is not knowledge. A prompt is a workaround. It is a set of instructions that compensates for what a specific model, at a specific moment, could not do on its own. Step-by-step scaffolding compensated for weak reasoning. "Verify your work" compensated for sloppiness. Elaborate role-play framing compensated for models that lost the thread halfway through.
Every model generation absorbs some of those weaknesses. Which means every model generation silently invalidates some of your workarounds. The prompt library, the step-by-step SOPs, the "proven templates" your team laminated and shared in Slack, all of it is inventory with an expiry date printed in invisible ink.
Most companies treat that inventory as an appreciating asset. They audit it, document it, train new hires on it. They are maintaining a depreciating one.
This is the twin of something I have written before. In Prompts Are Dead I argued that prompt engineering was never the skill worth mastering. This is the operational proof: the prompts don't just underperform, they actively expire.
What Actually Compounds
So if the prompts expire, what holds its value? In my own work, three things have survived every model transition. None of them are prompts. All three are leadership skills wearing a technical costume.
1. Knowing Which Model Does Which Job
I don't use one model for everything. I use the top-tier model, Fable 5, for planning and thinking. It is a genuine thinking partner. I bring it a problem, we work out what should be built and how. Then I hand execution to a cheaper, faster model to do the work quickly.
The expensive model designs the work. The lesser model does the work. That division of labor is a judgment call, not a prompt, and it carries forward no matter what ships next quarter.
Same with effort settings. The new models let you dial reasoning effort up or down. Medium is my default, and for thinking-partner conversations I often go lower. Most teams never touch the dial and overpay for every conversation they have.
2. Giving the Model Problems, Not Procedures
The old instinct was to write procedures: do this, then this, then this. The new models do their best work when you give them a goal, a few examples, and room to move.
Instead of dictating steps, I describe the problem I am trying to solve. I will ask the model to draw on best-in-class products, or to apply a real business framework. Jobs to Be Done, the framework Clayton Christensen taught at Harvard Business School, is one I reach for often. Given that kind of direction and the freedom to work, the model regularly comes back with an approach smarter than the procedure I would have written.
Notice what is durable here. The framework is decades old. The habit of framing a problem clearly is older than that. Neither one expires when a model updates. Your procedures do. This is exactly why we teach leaders to stack proven frameworks instead of collecting prompts.
3. Spending Your Time on Planning, Not Prompting
Here is the shift that surprised me most. I spend almost no time on prompting anymore. The hours I used to pour into engineering instructions now go into thinking deeply about what I actually want done.
That is the real skill behind using higher-end models well. Use the best model to help you plan. Think hard about the outcome, the constraints, what good looks like. Once the plan is sharp, a lesser model can execute it fast and cheap.
Vague goal plus perfect prompt loses to sharp goal plus no prompt at all. Every time.
Why This Is a Leadership Problem, Not a Technical One
Follow this to its end, because it changes who becomes valuable on your team.
In 2024, companies raced to hire and anoint prompt engineers. The pitch was that prompting was the new literacy, and the people who mastered it would be indispensable. But prompting skill is exactly the thing each model generation absorbs. The models keep getting better at understanding plain intent, which means the gap prompt engineering filled keeps shrinking. Betting your org on it is betting on a depreciating skill.
What actually appreciates are the three habits above: judgment about which model fits which job, the ability to frame a business problem clearly, and the discipline to plan before you delegate. Those get more valuable with every model release, because every release gives good judgment more leverage.
This is the difference between using AI and leading with AI. Using AI is typing a clever prompt into a box. Leading with AI is deciding what should be built, choosing the right model to build it, and knowing what good looks like when it comes back. The first skill expires on the vendor's release schedule. The second compounds on yours.
What to Do This Quarter
Three moves, none of them expensive:
- Audit your prompt library against the current models. Anything with step-by-step verification scaffolding is a candidate for deletion, not preservation. Shorter will often work better.
- Treat every new model generation as a breaking change. When one ships, re-test your critical AI workflows the way you would re-test code after a major dependency upgrade.
- Rebalance what your team practices. Less prompt-craft. More problem-framing, model selection, and planning. If you don't know where your team should even be running these models, start with the three levels of AI model usage.
The companies that win with AI will not be the ones with the biggest prompt libraries. They will be the ones whose people know what they want, know which model to hand it to, and re-tune fastest when the ground moves.
That judgment is exactly what we certify. If you want to know which of your people actually have it, that is what AI Officer certification is for.