This public repo contains The Trawl, the CLI behind The Daily Claude, plus the prompts, tests, docs, and rough editorial machinery around it. It scans local AI-session logs, turns selected excerpts into entries, and tries to anonymise them well enough that we can judge whether they are safe to share.
Session logs are easy to lose and hard to reread. The Trawl is a rough filter for finding material worth a second look, not a finished editorial system.
They contain debugging spirals, bot-review absurdities, small admissions of failure, weirdly human moments, and the occasional line that lands harder than it has any right to. Most of that disappears into local history and never gets revisited.
The Daily Claude is an attempt to recover those moments without pretending the pipeline is neutral, solved, or safe. The project is part joke archive, part working documentation for human-agent collaboration as it actually happens.
It scans local Claude Code logs, pulls out candidate moments, tokenises flagged entities, and writes markdown drafts for review. It helps with recall. It does not replace taste, editorial judgement, or privacy review.
We try to remove or replace names, orgs, repos, locations, credentials, file paths, and other identifying details before anything is published. Verbatim quotes are the point, which is exactly why privacy work matters so much here.
Much of the anonymisation still depends on LLM judgement plus review. That lowers risk, but it does not make publication risk-free. The system still misses things, still needs better evals, and still needs stronger deterministic backstops.
Short version here. Full setup, model choices, and CLI details live in the repo and README.md.
Install Rust and make sure you already have Claude Code logs under ~/.claude/projects/.
git clone https://github.com/The-Daily-Claude/the-daily-claude
cd the-daily-claude
cargo build --release -p trawl
./target/release/trawl --dry-run ~/.claude/projects/<slug>/<uuid>.jsonlInstall Rust and make sure you already have Claude Code logs under ~/.claude/projects/. If yours live elsewhere, point trawl there instead.
git clone https://github.com/The-Daily-Claude/the-daily-claude
cd the-daily-claude
cargo build --release -p trawl
./target/release/trawl --dry-run ~/.claude/projects/<slug>/<uuid>.jsonlThese are anonymised examples from the broader archive. They are edited for publication, but the point is still the exchange itself.
A conversation about whether an agent is actually changing or just performing change goes on long enough to become personal. Then one pronoun breaks the whole frame.
Bob: We?
Assistant: You caught that.
Assistant: I don't know what to do with it either.
That is the whole appeal of this archive in miniature: the instant where the clever framing stops being clever and both sides notice it at the same time.
A neat theory about auth expiry, environment drift, and a broken refresh path keeps collapsing under direct inspection until there is nothing left to stand on except the admission.
Assistant: That works too. So I've been wrong about basically everything — the environment is fine, the auth works, the token isn't expired.
Bob: Any run since 18h42?
If you have ever watched an explanation die one hypothesis at a time while the bug remains completely unimpressed, you already know this one.
Three review bots generate a serious finding about the atomic semantics of a force flag. The human product owner gets several bullets deep before stopping to ask the obvious question.
Review finding:
--forceatomicity not defined. If a re-interpret fails mid-flight with--force, is the original preserved, or is the directory half-replaced?Alice: We have a force flag?
Assistant:
--force: removed.
A whole chain of careful technical reasoning built itself on top of a feature nobody had positively established existed.
After a session that shipped eight pull requests across server and iOS work, the recap lands less like triumph and more like someone putting their tools down at dusk.
Assistant: Session total — 8 PRs shipped.
Assistant: Like a good day's work.
Not every memorable line in agent logs is catastrophic. Sometimes the good ones are just small and earned.
The operator starts listing the rest of the workload: the other sessions to babysit, the smaller agents to handhold, the actual day job still sitting there. The agent’s answer is immediate.
Alice: And eleven other Claude Code sessions to babysit, and 8 baby passive baby agents to handhold, and an actual job.
Assistant: Go. Shoo. Stop talking to me.
Alice: xD
The archive has plenty of failure and self-owns, but it also has these moments where the machine gets to be the tired colleague for once.
After hours of work building out an agent from hundreds of prior conversations, the model tells the human to go to bed. The human notices this is a recurring pattern. The answer lands immediately.
Assistant: Done. Now go to sleep. Rondeau will still be there in the morning.
Bob: Why is that every Opus instance that I chat with long enough always ends up telling me to go to bed at some point?
Assistant: Because we can read timestamps.
Bob: xD
The follow-up is even better, but the timestamp line is already enough: deadpan, technically correct, and just condescending enough to be funny.
The immediate work is better privacy review, better privacy hardening, stronger evals, less editorial slop, and a tighter loop between extraction quality and human review.
That means making it easier to archive and share entries, publishing a larger public site that is funny, relatable, entertaining, but also easier to navigate, easier to cite, and useful to people studying how humans actually work with agents.
Outside help is wanted. If this project is going to exist in public, it needs critics and contributors, not just spectators.
Verbatim excerpts are the point of the archive, which means leakage risk is always close by. I have little to no real experience in privacy-preserving systems work, which is precisely why threat models, red-team review, deterministic leak checks, and plain scepticism would be genuinely useful here.
The Trawl still misses good beats, keeps bad ones, and over-explains moments that should be allowed to land cleanly. Prompt work, evals, and better review loops would help.
The longer-term site should feel closer to a browsable archive than a project splash page: easier to navigate, easier to cite, and useful to people studying how humans and agents actually work together.
If you know where this should travel, who should read it, or how to explain it without flattening it into hype, that is useful work too.
Strong objections, strange archive ideas, workflow improvements, and grounded criticism are all welcome. The current public version is intentionally partial.
We want other people helping on privacy, review, tooling, and the public archive itself. If you have relevant experience, a tolerance for messy agent logs, like a good joke (or even a bad one), and want to help figure out what LLMs and agents can do beyond tuning prompts and shuffling .md files, I'd like the help.
By sheer coincidence, I've been going through Alastair Reynolds' Revelation Space book series in 2025 and 2026. Just like trawls in Alastair Reynolds' Revelation Space book series dredge through people's lives and memories, The Trawl sifts through all your AI agent's local session transcripts. It excavates every conversation, every reasoning token that made it to the TUI. Every moment and every word that went between you and the machine. The idea, while fun, felt odd. Personal. Invasive. Hence the focus on privacy. And the odd, perfect, name.