A research log Claude writes as it goes

A long research session produces forty findings and one chat window. A week later you remember that something was established about pricing and you cannot find which message it was in, or what the source was.

Have the agent write each finding as a bullet with its source in the note, in your own outline, so the research survives the conversation that produced it.

Set it up

The shape that makes it re-readable

One finding per bullet, the claim on the line and the source in the note, with the date it was read. That last part is what turns a research log into something you can still trust in March: a claim with no date is a claim about a day nobody can name.

- Workflowy’s MCP server cannot read or write notes
  note: workflowy.com/help/mcp-server/, read 2026-09-22.
        Seven tools, none of them touches the note field.
        Their own beta thread says the same.

Tag the ones you are unsure about and the ones you verified twice, so a later pass can tell them apart without re-reading everything.

Asking it back

Search reaches the same branch the agent wrote into, so the question in March is a search rather than a scroll. And a second agent, in a different client, reads the same branch: the research does not belong to the assistant that did it.

What it is not

It is not a transcript, and a transcript is what you get by default. The difference is that somebody has to decide, at the moment of writing, what is worth keeping. That decision is the whole value, and it is also why the agent should be told to write few bullets rather than many.

Deeper

• The reading notes nobody reads, and how to save them
• Never re-read what you have already read

Connect an agent, about two minutes · The twelve tools · Which note apps an agent can reach · Who runs this