Asking Questions of Your Own Saved Library Instead of Searching It
Search gives you a list. Sometimes what you want is an answer built from things you already read and trusted.
Commands over your own saves
Markwise's Copilot takes slash commands that run against your library and nothing else.
/find pulls things up. /ask gives you an answer synthesised from your saves. /note drafts something from a set of them. /space groups them into a collection. /save files a link without leaving the conversation.
The retrieval runs against your library only, so the answer cannot drift into whatever the model absorbed during training. Answers cite which of your saves they came from, which is the part that makes the output checkable rather than merely plausible. If a claim looks wrong, you click through to the source you saved and see for yourself.
The bug that taught me how this should work
An engineering detail I got wrong the first time, in case it saves someone else the debugging.
I was passing the user's question and the skill instructions to the search engine as one string. The search backend took that whole string as the query. So /find oil related bookmarks retrieved documents about bookmark managers, because "bookmark" and "find" and "related" are common words in a library about saving things, and they drowned out the one word that carried the actual intent.
It got worse with longer skill prompts. A few hundred words of instruction prepended to a three-word query means the query is 2% of the signal. The engine dutifully matched the instructions.
The fix was to separate retrieval from instruction entirely. Strip the meta-words from the query, so "find oil related bookmarks" becomes "oil". Retrieve on that real topic. Then hand the clean results, and separately the instructions, to the model. Two channels instead of one.
There is a general lesson buried in there. When you wire an LLM to a search index, be explicit about which string is the query and which string is the prompt. They look interchangeable in the code and they are not. Anything you concatenate into the query becomes vocabulary the retriever has to rank against.
Whether it beats plain search
I am not fully convinced synthesis beats a plain list for every query, and I built the thing. For "show me what I saved about X", a list is faster and more honest. For "what did those five posts actually agree on", a list makes you do the reading twice.
If you use AI over your own notes or saves, does it actually beat plain search for you?
The Copilot is part of Markwise, which is free to start.