Neural Gains Weekly Archive

An AI agent can hit every number in an instruction and still fail the task, because a content filter blocks what a model says, not what it is trying to do. This piece shows how to write goals a system cannot quietly game.

A prompt you re-paste every week belongs in an agent, not your chat history. Microsoft 365 Copilot Agent Builder loads the role, rules, and knowledge sources once, so each week you hand it only what changed, and the same pattern runs in Claude, ChatGPT, and Gemini.

A 4K screenshot does not reach the model at 4K. It is scaled to a resolution cap, cut into squares, and converted into visual tokens, which is why small text and exact numbers get lost, and why cropping tight before you upload changes the answer you get back.

A model does not read text. It processes tokens, and images, scans, and screenshots become tokens too, which is why you can hand it a photo of a page instead of retyping the whole thing.

Copilot, Claude, and NotebookLM all search by matching meaning instead of keywords, turning your files into a map of related ideas. That same shortcut can also return the nearest answer instead of the correct one, especially with messy data, mixed-up jargon, or oversharing permissions.

Every AI agent has a dial that decides how much it does before it stops and asks you. Here's how to find it, and how to brief an agent so its work is actually usable.

One of Claude's most useful features stays invisible until you flip a single setting, and it is the one that changes long-form work the most. Turn on Code execution and file creation and your writing becomes a workspace you edit in place and iterate on.

Claude seems to remember you on Monday morning, but that continuity is two separate systems, not one: a running summary of who you are, now free, and the paid ability to search a specific past chat. Knowing which is which is the whole game.