|
PRO / DEEP DIVE / ISSUE 01
The commitment ledger that won't let you drop a promise again
For seventeen years I have been losing tiny units of trust with my team and I never quite knew where it was going.
It went like this. Tuesday's leadership sync. Someone raises a hiring gap. I say "I'll get with Tara about that this week." I mean it. The moment is real. By Thursday afternoon I have shipped two product decisions, three customer fires, a board ask, and the hiring conversation is gone. Tara hasn't heard from me. She doesn't push. Next Tuesday she asks gently if there's an update and I'm caught flat-footed. "Sorry, slipped through."
Multiply that by twenty-six meetings a week and the math gets ugly fast. Not catastrophic, never the kind of thing anyone calls you out on. Just a slow drip of "Andrew said he would and then didn't" that adds up across the year into a real tax on your authority.
I tried every system. Notion databases with relations and views. A spiral notebook on the corner of the desk. An upgraded notebook with a fountain pen I bought specifically to make the ritual feel important. Every task manager from OmniFocus to Things to Asana. Getting Things Done twice through. None of it survived three busy weeks.
The thing that finally worked has zero discipline cost and runs itself on a Friday afternoon while I'm doing something else. Today's deep dive is how that ledger is actually built; what runs where; why every layer is what it is; and the four trade-offs you should know about before you commit to it.
> THE ARCHITECTURE
How it actually works
The ledger is the smallest possible system that can survive your worst week. Three components, all of them files on your laptop, none of them a database.
The first component is the corpus. Every meeting you have ever recorded with Granola sits as a markdown file in ~/meetings/ on your machine. One file per meeting, with the date and the attendees baked into the filename. This is your raw substrate. The harness reads it; you do not. You will probably never open one of these files manually.
The second component is the state file. A single markdown file at ~/meetings/commitment-ledger.md that the harness owns and rewrites every time it runs. Two big sections. Commitments I made. Commitments others made to me. Each line carries the date, the person, the source meeting filename, and a status of open, done, or overdue. A third section appears at the top when the harness detects items that have been open more than fourteen days. That section is called Slipping and it is the single most important UX choice in the whole system.
The third component is the prompt. One pasted block of text that tells Claude what to read, what to look for, what to merge with the existing ledger, and where to write the output. You run it manually for the first two weeks. By week three you wrap it in a Claude Routine and it fires itself every Friday at four in the afternoon. By month two you stop thinking about it the way you stop thinking about your calendar invites; it's just there.
The reason all three components are markdown and not a database is that I want the ledger to outlive every tool in this stack. Granola might disappear. Claude Code might be replaced. The Routines feature might get renamed. None of that touches the actual ledger. It's a text file. I can read it in Vim, I can pipe it into the next model that ships, I can email it to myself, I can hand it to a new assistant and they understand it in thirty seconds. The format is the moat, not the tool.
There is one more design choice worth naming. The harness never silently overwrites a commitment as done. It can only move a line from open to done if it found language in a later meeting transcript that confirms the close. "I sent Tara the JD on Wednesday" gets the line closed. The harness's default when it isn't sure is to leave a line in open with a confidence flag. This costs you a little manual review on the Friday refresh, but the alternative (Claude marking things done that aren't actually done) is the failure mode that would kill your trust in the system within a week. Graceful degradation beats false confidence. Always.
> THE PRODUCTION PROMPT
Paste-ready
This is the prompt I run today. The version in the public workflow article is the eighty percent version; this is the version that has been through the trade-offs.
You're maintaining a commitment ledger for me from the meetings in ~/meetings/.
Read every transcript file in ~/meetings/ from the last 7 days. Also read
the existing ledger at ~/meetings/commitment-ledger.md if it exists.
Produce a single updated ledger with three sections, in this order:
> TRADE-OFFS
What this sacrifices
There are four. Pretending there aren't any is how you end up with a system you eventually rage-quit.
The privacy trade-off. Everything lives on your laptop. The transcripts, the ledger, the prompt history. Claude reads files from your machine and writes back to your machine. Nothing ever lands on Notion or a vendor's S3 bucket. That is enormous for sensitive commitments (hiring decisions, comp discussions, board context). The cost: if your laptop dies, your ledger dies. The mitigation is Time Machine plus a manual Friday-afternoon git commit of the meetings folder to a private repo. I do both. The full SaaS-style version of this workflow would put the ledger in a vendor's cloud and give you "anywhere access," and you would also be one breach away from every commitment you ever made being readable by someone else. Pick your fear.
The classification trade-off. Claude is good at picking commitments out of conversation. It is not perfect. About one in every twenty entries on a busy week is a misclassification, and about one in every fifty is a hallucinated commitment (someone said "we should probably look at that" and Claude reads it as a commitment to look at it). You have two options for handling this. Option one is to trust the system and fix the wrong ones as they surface in real life ("hey, I never actually said I'd do that"). Option two is the Friday-afternoon ten-minute review where you eyeball the new lines and delete the bogus ones before the ledger is canonical. I do option two for the first month with every new harness, then move to option one once I've calibrated my own trust.
The 1:1 dynamic trade-off. Walking into a 1:1 with the ledger open changes the conversation. Most of the time it changes it for the better; the person across the table sees you have receipts and they relax because nothing is going to be lost. About one time in fifteen it changes it for the worse, when someone reads "I have receipts" as "I'm being audited." You'll know in the first three meetings which of your direct reports is which. For the second type, leave the ledger closed during the meeting and use it before and after. The information is still working for you; it just isn't sitting visible on the desk.
The discipline tax. The harness writes the ledger every Friday whether you read it or not. If you don't open it, it just keeps growing. After three weeks of not opening it, the Slipping section is depressing rather than useful. The hard truth is the ledger only works if you read it. The fifteen-minute Friday block is non-negotiable. If you cannot commit to that fifteen minutes, do not start the ledger. Pick a different workflow this week.
> THE MOVE
What to do next
Tonight, fifteen minutes. Open Claude Code from inside your ~/meetings/ folder. Paste the production prompt above. Run it. Open the resulting commitment-ledger.md. Read the Slipping section first. Pick three things. Send three messages tonight closing them out or rescheduling them honestly.
Tomorrow morning, three minutes. Put a fifteen-minute recurring block on every Friday at four in the afternoon. Title it Refresh ledger, close what I can. This is the ritual. Miss it twice and you start losing the trust I described in the lede.
Two Tuesdays from now. Wrap the prompt in a Claude Routine and trigger it from a cron at Friday four in the afternoon. The Routine writes the ledger before you even sit down with it. Now the only discipline cost is reading.
Next month. Bring the ledger to your weekly leadership sync. Not as an audit. As a posture. "Three things from last week are still open. Here they are. Tara, you and I; me and the board; me and you." This is what the top one percent of operators look like. Now you do too.
|