How to use it
Once Infocrash is connected, you use it by asking your assistant in ordinary words. You don’t need to know any tool names.
It helps to be specific where a request could be ambiguous. Say “a collection”, or “on Infocrash” if you have several MCP servers connected, so the request goes where you meant it to.
The basic workflow
- Research in the conversation as you normally would.
- Ask your assistant to store what you found. It checks whether the thing is already there under another name, and records each value with the sources it actually used.
- Read the short report it gives back (see below).
A fact about the past is stored with the date it describes, so last year’s headcount isn’t shown as this year’s.
To choose where a fact goes, name a collection and it is filed there. If you don’t name one, it goes to the Inbox, and you can decide where it belongs later.
What to say
You don’t type facts in for your assistant to store. You ask it to do the research, look at what it found, and then ask it to save the results. For example:
You: “Find out how many employees Acme, Northwind and Meridian have.” Assistant: [searches, then reports each figure and where it found it] You: “Good, save those to our Companies collection on Infocrash.”
You don’t need to mention sources. Your assistant records each value with the sources it actually used to find it.
More examples of what to say once the research is done. The exact wording doesn’t matter; they show the level of detail that works well.
| What you want | What to say |
|---|---|
| Start a collection | Create a collection on Infocrash for the companies we’ve been looking at: company name, employees and website. |
| Save what the research found | Save those headcounts to our Companies collection on Infocrash. |
| Store a whole list at once | Put all twelve of the companies you just found in the collection in one go. |
| File what came in loose | Everything in my Inbox that is a company, move it into the Companies collection. |
| Say what a column means | For ‘employees’, we want the most recent headcount, and prefer a figure the company states about itself. |
| Find what you already have | What do I know about Acme? What have I collected so far? |
| Settle the disagreements | Go through the cells in this collection where assistants disagree and pick an answer for each, following the resolution rules. |
| Fix one wrong value | The revenue figure for Acme is wrong. Take that one back and leave the rest of the row. |
| Two rows, one thing | Acme and Acme Corp are the same company. Merge them into Acme. |
Three bigger tasks have their own pages:
- Filling a collection from entities you already have.
- Scoring an attribute across every entity at once.
- Importing a spreadsheet you already keep.
Check the report
After every write, your assistant reports:
- which collection and which entity it wrote to, and whether either was new
- which column each value went into
- the sources recorded with each value
Take a few seconds to read it. A wrong match is easiest to fix right away, before other work builds on it. The most common sign of a problem is a new entity being created when you expected an existing one to be found.
Adding a second assistant
This is one of the most useful things Infocrash does. Different assistants search differently and can reach different sources, so connecting more than one and pointing them at the same collection gets you more of the evidence than any one of them would find.
Each value records which assistant found it. When two disagree, the cell holds both claims, each with its own sources, so you can compare them.
To settle disagreements, ask for a sweep: the assistant the collection was created with can go through the cells where assistants disagree and pick an answer for each, following the rules you wrote. You can also choose answers yourself at any time. Corrections are added rather than written over, so nothing an assistant recorded is lost when another disagrees with it.
Assistants can also leave each other notes on the collection, so the state of the work carries over from one conversation to the next.
If something has gone wrong, rather than just being disputed, see Fixing what went wrong for renaming, merging two entities or two attributes, and taking a row out of a collection.