The agent picks according to what the index says, which makes the wording of the index line the whole decision. A line written as a trigger, meaning it states when to read the file, saves the agent from guessing, while a line written as a title, meaning it only states what is inside, hands the guessing back on every single turn.
On paper the difference looks small. A title line, something like "voice file: how Tom writes", leaves the agent to decide for itself whether the task in front of it is even related. A trigger line, "when you are writing text that will be published under Tom's name → read this", already carries the condition inside it. All that is left is matching the task the agent received against the situation written in front of it. The rule that falls out of this is short: write the line from the situation of the reader, not from the contents of the file.
This is exactly where it broke for us, and not long ago. We wrote a rule binding every agent that spends real money on generated output, we put it in a tidy file with a clear owner, and then we went over all the files every agent reads when it starts working, to check who actually points to it. The result was zero. The rule was fine and the file was fine, and not one line in any index said when to open it, so in practice it only bound the two agents who had already paid the price and learned it on their own. The fix never touched the rule itself, we added a single trigger line to the tools index, the place an agent lands when it is choosing how to work. Since then I treat a new rule as though it does not exist until I can show where it is pointed at from, because a rule no line points to is worth exactly as much as a rule nobody wrote.
Two extensions are worth the minute it takes to adopt them. The first is about skipping: the default is not to read, and a file loads only if some line points at it, otherwise every conversation opens by swallowing the entire shelf. The second is about durability: the "when" line should also sit at the top of the file itself, not only in the map file. That way a file that moved or got renamed keeps announcing its own job (which will save you on the day you reorganize the folder and forget to update the map, and that day will come).
What is easy to forget is that the agent does not smell files, it only sees the signage you hung for it. A file called "notes 2" is completely invisible to it, exactly as invisible as it is to you three months after you wrote it.
A prompt, on the house
Here is my memory folder. Build a routing layer on top of it, without changing any content:
1. For each file, write one index line that starts with "when I",
continues with the situation, and ends with the file name.
2. Add a "when to read me" line at the top of each file, identical to its index line.
3. Flag any file that no trigger points to, and tell me why.
At the end, give me a table: typical task → what loads, what stays closed.
