The task shipped. Weeks later someone asks how it was done, what was tried, why the other option lost — and there is nothing. The knowledge existed for exactly as long as the work did. Record-keeping always loses to the next deadline.
Why this keeps happening.
Recording has always been separate work done after the real work — and anything scheduled "after" loses. The fix is not discipline; people have tried discipline for decades. The fix is making the record a byproduct: something the working session produces on its way to finishing, not a second task.
The way of working.
Sessions log as they work
An Agent session keeps a structured map — decisions, open questions, completed steps — updated at the moment each thing happens. When the session ends, the record already exists.
A completion checklist routes the byproducts
Every finished task answers a few standing questions: does this become a test? a reusable pattern? a public update? a documented way of working? Whatever qualifies gets filed — by the Agent, at completion time.
Experiments get a canonical record
Anything with a hypothesis and an outcome is written to a standing experiments log, so results accumulate instead of evaporating with the chat that produced them.
Our workspace has a standing rule: at every task completion the session answers four questions — test? recurring problem? release channel? way-of-working candidate? This article exists because of that rule: the working method it describes was filed as a content candidate by the session that used it.
Start here.
The core workflow shows how sessions record as they go. Run one real task through it and look at what is left afterward.