Your Claude Code sessions are your best data, and by default they’re deleted after 30 days

Claude Code writes every ses­sion to a plain­text file under ~/.claude/projects/. By default it keeps them for 30 days, and then a back­ground sweep deletes them without showing a message.

I think those files are the best data I have about my own work. They record what I asked for, how I asked, where I cor­rected the model and what I gave up on. Nobody writes that down by hand.

One session at a time, and across sessions

At the end of a work­ing ses­sion I already have Claude write a short reflec­tion: what I cor­rected, which rule is worth keeping. That works one ses­sion at a time, and it can’t see a pat­tern that only shows up across sessions.

Reading all the tran­scripts can. One read found about 21 ses­sions where I had asked, in dif­fer­ent words each time, for the same thing: where does this ses­sion stand, what is closed, what is still open. Now it’s a com­mand, /recap. My work­ing rule since: a request you repeat is a skill wait­ing to be written.

What the record shows

I typed roughly 11,500 mes­sages from March to October. Nearly a quar­ter went to Alisa.ua, the online store I built, mostly from March to May. Building my own prod­ucts took about 14%, and Claude Code itself about 13%: set­tings, skills, the machin­ery around it. Media and ad cre­ative took about 12%.

What the existing tools miss

I also read the source of five open-source tools that parse these files: cc-transcript, introspect, claude-code-analytics, claude-hindsight and Simon Willison’s claude-code-transcripts. Each is built for its own job: HTML pages, search, dashboards.

None of the five sorts ses­sions by who started them.

What was said lives in three places, not one. Most of it is in ordi­nary user lines. Lines marked “meta” hold the mes­sages another ses­sion sent this one, which you want, and every skill file Claude loaded, which you didn’t write. Four of the five ignore that mark, so a meta line looks like some­thing you typed. And a prompt you type while Claude is still work­ing is stored as a dif­fer­ent kind of line, which only cc-transcript reads.

And none of the five removes secrets.

The files hold secrets

A scan of mine found dozens of strings shaped like API keys. Almost none were in any­thing I typed; they came in with the work, through files the model read and com­mands it ran. Some were keys I still use.

What I’d do with yours

• I’d raise cleanupPeriodDays in settings.json. Setting it to 0 fails val­i­da­tion, and the docs sug­gest a large value such as 3650. I set 3650.
• I’d back the folder up off the machine, encrypted, and never into git.
• I’d parse it once into a small corpus and dig into it, instead of the raw files. Mine keeps the facts for each ses­sion and the text of each mes­sage, drops tool results, redacts any­thing shaped like a cre­den­tial, tags each ses­sion by who started it, and keeps its rows after Claude Code deletes the transcript.

What I’d mine it for

• Requests I repeat. Each one is a skill or a com­mand wait­ing to be written. That is where /recap came from.
• Corrections I repeat. A cor­rec­tion that turns up in sev­eral ses­sions is a rule for CLAUDE.md or a skill, not some­thing to say a fourth time.
• Questions I keep asking. About half my ses­sions about Claude Code itself were three mes­sages or fewer, mostly a quick “how do I…”. The answers are worth a note I can find again.
• Permissions I keep approving. Commands I allow by hand every day are allow rules wait­ing to be written. Claude Code comes with a /fewer-permission-prompts skill that does this for the read-only ones.
• Kinds of work that recur. Sorted by kind, my ses­sions sug­gested which jobs could go to an agent of their own: nine can­di­date roles, from an ads oper­a­tor to the Alisa.ua work.
• Secrets that came in with the work. Anything shaped like a key, so the live ones can be rotated.