I've been experimenting with various Claude skills and patterns long enough (6+ months) to see which ones actually stick. One interesting concept has emerged from this: work logs.

A work log is a detailed summary of the work done during a development session. I use a journal-like format with daily entries, as one working day is a natural unit of time to summarize. Claude writes it based on all the sessions we've worked through during the day. It's quite detailed and includes not just a technical summary but also notes from our discussions, what was contemplated and why.

The trick to compiling this is to capture it from all the claude sessions throughout the day. This is why the process consists of two separate skills:

/snapshot-session writes a brief summary of one session. I typically /clear the context often, easily a dozen times a day, and take the snapshot just before clearing it. It's fairly quick to run, about a minute or two. Each snapshot goes into a temporary file.

/end-session compiles all snapshots into one daily journal file. It takes quite a bit longer, so it's the last thing I do at the end of the day, just leaving it running. This also cleans up the snapshot files.

The output goes into a markdown structure that looks like this:

work-log/
 ├── 2026-05/
 └── 2026-06/
 ├── 01.md
 ├── 02.md
 └── …

This becomes a knowledge base that accumulates over time. It's not immediately useful, but as the entries pile up the value becomes more obvious.

Often you come across a situation where your client or colleague asks something and you think: "didn't we talk about this a couple of months ago?" or "I'm pretty sure we've already covered this." That's when you can send claude searching through your knowledge base.

This setup becomes even more useful when you integrate external channels like Gmail or Slack MCP into your claude project. When someone emails you about a bug or feature you don't copy-paste it into the work session, but instead send claude to read it.

This brings additional information into the context: the name of the sender or group, the people involved. claude loads these into the context window and they get baked into the snapshots and work logs. Over time your coding agent learns to answer things like "someone asked about this before but I don't remember who it was."

It takes a little effort to set up the skills and get into the habit of using them, but that's the most critical part: you need to capture the information when it's fresh. This has proven useful in just a few months, and it's easy to see what it'll do over a few years. But you can't recreate a work log after the fact, the day's already gone.