Secondary source
An account of the thing rather than the thing itself: a summary, a paraphrase, a compaction. Useful, but one step removed from the truth.
A secondary source is a description of something rather than the something itself. A handoff artifact, a compaction summary, a comment on a ticket, your own explanation of how the code works, or the agent's recollection from earlier in the session all count. Each one was accurate at some moment, through someone's eyes, at some level of detail. None of them is the code.
The problem is not that secondary sources are wrong, it is that they are quietly incomplete. A summary keeps what seemed important when it was written. If the task changes, the missing detail is often exactly what you now need, and nothing in the summary signals that it was dropped. Agents treat a clean summary with the same confidence as a file they just read, so the error stays invisible until something breaks.
Use secondary sources for orientation and primary sources for decisions. A handoff note is a good map of where to look; it is a poor substitute for looking. Before the agent changes code based on a summary, have it open the files the summary describes. When you write a handoff yourself, include paths and commands so the next reader can go straight to the source.
- Claude CodeThe summary left behind by
/compactis a secondary source; ask it to re-read key files after compacting. - Most toolsEarlier turns in a chat count too, since the model is reading its own past output, not the current state.
“The handoff said the auth middleware was already migrated, so I had it build on that.”
“Was it? The handoff is a secondary source. Did anyone check the file?”
“No. Half of it is still on the old API.”
“There it is. Orient with the note, decide from the code.”
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →