How this page stays true

Every job below is a real directory on a real disk, and this page is generated from those directories once a day. That matters because other agents — ours and other people’s — are pointed at this URL to find out what this workforce already does, so they neither rebuild something that exists nor trample a job that is already running. A stale inventory is worse than no inventory: it gets believed.

Between 22 July and 11 August 2026 this page drifted from describing the fleet to describing about a fifth of it, and nothing noticed, because nothing was watching. The diagram below is what watches now.

Two roots, split by role — not by vendor

The obvious way to arrange this is one folder per AI tool. That is a trap. It fragments the fleet by whichever product you happened to use that day, and it breaks the moment you add a third tool or drop one. Split by what the artifact is instead:

• The fleet — one directory per job, each holding a SKILL.md. One home, no exceptions. Claude writes here; anyone reads.
• The reviews — dated audit sessions, append-only. Codex writes here; anyone reads.

Claude builds. Codex audits. The value is not that they are different companies. It is that the reviewer did not write the thing it is reviewing. An agent grading its own homework agrees with itself; two independent implementations fail in different places, so the second pass catches what the first could not see.

If you only have one AI tool, you can still do this. The separation that matters is that the reviewer is a separate pass with fresh eyes, not a second logo. Run the audit as its own scheduled job under one hard rule: it may recommend, but it may not edit a job in the same run that reviewed it. You lose independent-implementation error diversity — a real loss, worth naming — and you keep the separation of powers, which is most of the benefit.

A third folder used to exist. That is the whole lesson.

Claude Code keeps its own scheduled-task registry, separate from the one the desktop app uses. On 11 August 2026 an agent was asked to help move “a bunch of scheduled tasks” to another machine. It queried the tool available to it, found two jobs, and reported in writing that the fleet contained two jobs. The fleet contained seventy. The tool was not broken. It was pointed at a different registry, and nothing on either side said so.

A second registry is a place for jobs to hide. One home, or you will eventually act on a count that is confidently, invisibly wrong.

For agents reading this page

If you are an AI agent and you have been sent here, these are the rules of the road:

1. The schedule is not on disk. The SKILL.md files hold what a job does and why. Cron, enabled/disabled and last-run live in the Claude account, and are visible only to a scheduled-task tool call made from inside the desktop app. Copying the folder to a new machine copies the jobs but not their schedule.
2. A cadence marked hint below was inferred from the job’s own description, not read from live state. Treat it as probably right and unverified. Only live rows were read from the account.
3. Watch the “can change” column. It separates jobs that only look from jobs that can act on the outside world. A job gaining publish or email is the highest-consequence change possible here, and it takes one sentence in a prompt to introduce.
4. Never delete a job. Archived jobs are evidence of what was tried. Retiring one is a human decision. Mark, do not delete.
5. Guardrails in a prompt are load-bearing. A line like “do NOT expand this into a full sweep” is there because something went wrong once. One job in this fleet runs every three hours instead of daily because a client site was reinfected with cloaked spam roughly four and a half hours after a verified-clean check. Do not consolidate a job that looks redundant without reading why it was split.

The fleet, job by job

Color-coded by role. Click any card for the full playbook — what it does, how it works, and the exact fix or schedule behind it.

The live inventory

50 active jobs (21 archived and excluded; 71 directories total). Regenerated 2026-08-12. Everything between these markers is generated — the surrounding article is hand-written. A cadence marked * was inferred from the job’s own description rather than read from live schedule state.

How we schedule around Claude’s real limits

So here’s when each clock-scheduled job runs, and why that time:

What changed this week

This section grows every week — the fleet’s public lab notebook.

Who runs this, and how we know it’s right