Better instruction following can expose the mess in instructions I wrote myself. That is the Opus 4.7 detail I care about most: in its April 2026 announcement, Anthropic emphasized more literal instruction following and recommended attention to prompts when changing versions. There is plenty of maintenance work hidden in that warning.
In July 2026, I suspected that rules created to organize models were constraining newer models. In September, I expected an agent to research missing information before returning questions. Looking at this in October, I need to apply that expectation to my own text. Asking for initiative requires leaving room for it to happen.
A contradiction is easy to construct. In a hypothetical example, one rule says to investigate until the cause is found; another requires confirmation before opening any additional file. The agent reaches the next file and stops. Then I complain that it did no research. I wrote the obligation to stop and somehow still managed to be surprised by obedience.
I want to explain why each rule exists. An instruction added after a particular failure can remain long after the situation changes. Keeping it out of habit imposes a cost on every future task. Nearly identical versions of the same instruction deserve review too: different wording can look like different authority to whoever must execute it.
My migration evaluation would include two requests. In one, missing information is accessible through research. In the other, one of my decisions is missing. I want the instructions to distinguish these cases before execution starts, without an intervention halfway through to explain the secret meaning of a sentence I wrote.
When something stalls, adding another paragraph is tempting. I prefer locating the conflict first. Sometimes I described a mandatory sequence when only the expected result was necessary. Sometimes the rule sits far from the action it governs. A small edit makes behavior easier to compare; rewriting everything makes it harder to discover what helped.
I also want the agent to identify the instruction that prevented progress. That turns irritation into a concrete decision about its working contract. My next step after this announcement is to reread accumulated rules and mark those that depend on a generous interpretation. A more literal model has every right to take what I told it to do seriously.