Your Old AI Instructions Are Making the New Models Worse

Anthropic launched Claude Fable 5 in June, then Claude Opus 5 in July, delivering nearly all of Fable 5’s intelligence at half the price. Both models got capable enough that the team at Anthropic had to go back and rewrite how it talks to them.

They deleted over 80% of the instructions it gives its own AI coding tool and the models’ performance didn’t just hold steady they got better.

The team didn’t cut those instructions just because the model got more powerful and could afford to ignore extra guidance, but because the extra guidance was actively working against the new model.

Similar to how a company policy handbook fills up with contradictions nobody’s caught in years, because the person who added the second rule never read the first one. Anthropic’s original instructions might have said “leave documentation as appropriate.” while, a few lines below it, another said “never write more than one comment per function.” These two rules, added at different times for different problems, never reconciled with each other. Feeding contradictions like that to a model that’s already extremely capable of good judgment means it ends up chasing rules that fight each other instead of doing the work.

The problem gets worse over time, because instructions written once get applied to every future case, forever, whether they’re still relevant or not.

Picture handing a new cook two sets of instructions for the same dish. The old version is a 40-step list covering every possible mistake, exact gram counts, exact seconds on the stove. The new version is a recipe card: here’s the dish, here’s what it should taste like, use your judgment. A cook who actually knows how to cook does better with the recipe card.

If your business has instructions, prompts, or skills written for an earlier Claude model, check them before you blame the new model for anything that goes wrong. If Claude is running with its own editable setup, a CLAUDE.md, a config file, whatever you call it, copy the link to Anthropic’s writeup and paste it into your Claude, then ask it to update its own instructions to match how the new models actually want to be talked to.

At TigerTiger, we build and update that setup for the businesses we work with, so it stays matched to whichever Claude model you’re actually running instead of the one that existed when it was first written. See how it works.

Source: The new rules of context engineering for Claude 5 generation models