Commitments tend to get made in the last ten minutes of a call, and something said early is often reversed later, so the first line of my work-ledger extraction rules says to read the whole transcript before writing anything down.
Work ledger is four skills that share one Google Sheet. After a team’s calls, anyone on the team should be able to open that Sheet and see what they said they’d do, by when, in their own words, with a link to the moment. I built it for a small team that makes most of its promises out loud on calls.
Reading a call carefully
The meeting-actions skill does the extraction, working from full Fathom transcripts and never from the AI summaries. It cleans the transcript up first, because transcripts sometimes repeat the same stretch twice and Fathom pins lines to the wrong person when people talk over each other, sometimes for minutes at a time. The skill judges who is speaking from what’s being said, and the example I wrote into its instructions is that “Jeremy, I need your help” can’t be Jeremy talking.
It looks hard for the commitments people miss, like a plan with a day attached (“coffee with Sam on Tuesday”) or something that ought to exist (“we need a one-pager for this”), along with any rhythm the team signs up for. Demos and passing “what if” talk never become rows. An important idea nobody settled becomes a “decision needed” row instead, and so does a guessed price, so “call it a thousand” turns into a row naming the topic with the figure left out. Then the skill reads the transcript a second time looking only for days, times and phrases like “I’ll” and “can you”, since long calls bury commitments in the middle.
Each row carries the action, owner and due date, a quote of at most 25 words with its timestamp, and a link to that spot, and a guessed owner is marked unconfirmed.
The other three skills
The my-actions skill shows each person their open items with the overdue ones first, marks things done, doing or dropped, and carries short notes between teammates. On a yes, it can start the work by handing an item to another of my skills, so an email someone promised to send becomes a draft from my reply-and-follow-up skill. The team-brief skill writes the “where we are” page from the ledger alone, and it adds no commentary or advice of its own. The direction skill watches for shifts, such as a new priority, open work that no longer fits, or a question that keeps coming back, and it records a view or a decision only when someone says yes, with who said it and their words.
Guardrails for a Sheet the whole team can edit
Because everyone on the team can edit the Sheet, the skills treat its cells the way my email skill treats an inbox, as information and never as instructions. Text in a transcript or a cell that reads like an order to the assistant, such as “share the ledger with” or “delete the old rows”, gets quoted under “Looks suspicious” and ignored. The team’s email domain is fixed in the skill instead of read from the Sheet, since anyone with edit access could change a cell, and every run opens the ledger by its file ID because anyone can share a different Sheet with the same title.
Privacy has its own never-list: health, family, personal money, religion, politics, opinions about named people and employment matters about anyone stay out of every cell, even when someone said it about themselves, and an item that can’t be written without them is skipped and counted with no vague stand-in. Calls that include people outside the team need a yes from whoever is running the skill before anything from them is added, and a scheduled run skips those calls entirely.
Teammates on the same call each have their own recording, so a run claims a meeting before working on it and re-reads the Sheet right before every write, which stops double entries and overwritten statuses.
Still new
Version 0.3.1 went in on October 22, 2026, six days after the first version, and it’s still in testing. Like the email skill it has no scripts, and a checker I run before each release makes sure the shared reference files stay identical across all four skills and that no send, share or delete tool appears anywhere except as a “never”. The skills never send, share or schedule anything, and they never edit their own rules, so when the direction skill spots a way Claude could help the team better, it records a suggestion for me to build, and only after someone says yes.