The workaround outlived the bug

Opus 5 shipped yesterday. I skipped the benchmark table and opened the prompting guide. It told me to delete instructions I had spent months getting right.

Anthropic's own documentation asks you to strip verification instructions out of your prompts. "Include a final verification step." "Use a subagent to verify." On Opus 5 those lines cause over-verification: the model already checks its own work, so the instruction piles onto behaviour it already has and burns tokens for nothing.

I have those lines. I wrote them months ago, and back then they were correct.

Every instruction is a fossil of a failure

Read your own instruction file closely and you're looking at a scar record.

"Don't leave stubs" is there because a model left stubs. "Verify before you answer" is there because one answered with total confidence and got it wrong. Every rule in my AGENTS.md and my skills is a fossil of a specific failure, from a specific model, on a specific afternoon when I got annoyed enough to write it down.

The catch is that fossils don't announce when the animal comes back to life. The guide flags one more: if your code-review prompt says "only report high-severity issues" or "be conservative," Opus 5 takes it literally and reports less. The exact line I'd have written to stop an older model from drowning me in noise now suppresses real findings.

Which lines actually expire

Not every line rots.

Some lines encode a model's weakness. Others encode my decision. "Verify your work" is the first kind. "Hungarian is never a literal translation of the English" is the second kind, and no model release will make that obsolete. The first is scaffolding around a hole. Taste doesn't get patched.

So the audit is a single question per line: did I write this because the model couldn't, or because this is how I want it done? The first kind has an expiry date. I keep the Phora agent layer honest the same way, and the first kind is usually the larger pile.

The stack has technical debt too

We accept that code needs maintenance. Nobody says that about the instruction files sitting on top of the model, and they age faster than the code does, because the ground under them moves every few months.

Every workaround ships with an expiry date nobody writes down. Mine sat there for months after the bug it was written for had been fixed, making the work worse the whole time.

The model updates itself. Your picture of it doesn't.