What is the most efficient way to update an AI agent on recent project changes?
by•
I realized our system prompt covers our overall app structure but fails to inform the AI about work completed in the previous session.
Relying on raw git commit logs often gives too much noise while writing manual daily summaries feels like unnecessary overhead before starting work.
Do you use a lightweight change log file automated git diffs or a brief session recap to keep your agent aligned?
8 views
Replies
Filter the log to one file. In a competition repo I keep a context file and commit it on its own, with the current state as the message, like "training paused for the reboot". Then git log on that file alone reads like a session recap, without the noise of every code commit. The file says where things stand, and its log says how it got there.
In my experience the noise in a git log comes from how the commits are written, not from the log itself. "fix", "wip" and "update stuff" are noise. "Checkout kept warm: a cold start took 7 to 11 s" is a recap. Since the agent writes the commits anyway, the overhead is zero once the style is a rule in your project file, and a one line log of the last twenty commits is enough to start a session. Anything that needs more than a line, a decision or a vendor quirk, goes into a topic note right when it's learned. Enio's single file trick works well on top of that for the current state.
Raw git logs are too noisy for me too, and I did not want to write summaries by hand either. So it's all scripts and hooks.
A scheduled job reads the Claude Code conversation logs and writes a worklog for each project with what I asked for, which files changed, the main commands and anything that went to production. No LLM summarizing. It also archives the logs, since Claude Code deletes them after 30 days.
Then a Stop hook writes a small session card every time the agent finishes a turn: my last instruction, its last answer, a pointer to the full log. A SessionStart hook loads it. If a session hangs, the next one picks up right where it stopped.
And every project has a memory folder. CLAUDE.md tells the agent to write down when I correct it and why, anything it changed in production, and incidents with their causes, and only the index loads at the start, one line per memory.