Build a cross-meeting commitments ledger
Pair meeting minutes with a living **commitments ledger** context file so an agent flags **said-we'd-do** lines that vanish across weeks — with owners, stall reads, and next steps **you** own, grounded in quoted text so nothing is invented.
Plus: three prompts — reconcile **new minutes → ledger deltas** only, draft **one sendable follow-up** per OPEN row (no auto-send), and run a **priority drift** pass when your objectives file sits beside the last four weeks of notes.
| Audience | Everyone · Executives |
|---|---|
| Time | ~12 min |
| Prerequisites | You can create agents and attach context files ([Create an agent from scratch](/tutorials/create-an-agent-from-scratch), [Add your first context file](/tutorials/add-your-first-context-file)). You already extract action items from a single meeting ([Turn meeting notes into action items and decisions](/tutorials/turn-meeting-notes-into-action-items)) — this lesson is the **cross-meeting** ledger layer. Helpful: [Keep your context files honest and fresh](/tutorials/keep-your-context-files-honest-and-fresh) for refresh rhythm; [Set up your Monday morning briefing](/tutorials/set-up-your-monday-morning-briefing) if you want the same scan on a calendar. |
| You'll end up with | A **Commitments ledger** context file (Markdown table or bullets: commitment, owner, source meeting, status), a **Meeting follow-through** agent chartered to reconcile new minutes against that ledger without inventing tasks, and one completed reconciliation pass you can defend in standup. |
When a tutorial shows italic text in quotation marks, it usually mirrors a label or helper string inside Auxot. Product copy changes between releases — if something reads differently in your workspace, trust what you see on screen.
Callouts with a Worth knowing gold accent are meant as must-read context before you move on. Blockquotes that open with Tip are lighter, optional depth.
Why this matters
One meeting, cleaned up, still leaves a blind spot: commitments that sounded real once and never surfaced again. The weekly standup covers this week. Nobody has a job title that says re-read October’s strategy off-sites and compare them to what we actually shipped.
A context file (durable business memory the agent reads every time) plus a short reconciliation habit turns scattered minutes into an honest backlog: what we said, who owned it, whether it ever came back. The agent does the tedious comparison; you decide what is still worth the energy. Today you wire the ledger. The next time you ask what did we quietly drop?, the same agent answers in minutes instead of never.
The agent reads what you file. It does not attend meetings for you.
Quick start
- Create a context file — Settings → Context Files → Add: name it something you will open weekly (for example
Commitments ledger). Use a simple table or bullet pattern: Commitment | Owner | Source (meeting + date) | Status where status is OPEN, STUCK, or DONE only. - Seed the ledger once — paste three to six real commitments from recent notes (redact customer names). If you do not have clean minutes, paste rough bullets and label sources honestly. Empty rows are fine; invented history is not.
- Open Chat with the Admin Agent — click Chat in the left menu, ask it to create a Meeting follow-through agent whose only job is: read new meeting notes you paste, compare them to the ledger file, propose deltas (new rows, status changes, duplicates to merge), and refuse to invent commitments not grounded in quoted text.
- Attach both files to that agent — the ledger plus a Minutes inbox file (or attach per prompt if you prefer one-off pastes). Lock the rule: no quote, no new commitment.
- Run one reconciliation — paste last week’s notes, accept or edit the agent’s table patch, then archive the paste into your minutes file so next week’s diff is honest.
Done? The agent appears under Settings → Agents, the ledger file exists, and you have one reconciliation thread where at least one OPEN row has a named owner and a you-owned next step (even if that step is close as won’t-do).
The agent can do that?
1. Minutes-to-ledger delta with quotes only
Open chat with Meeting follow-through (or the Admin Agent if you are still iterating the charter):
Here are new meeting notes (below). Compare them to the Commitments ledger context file. Output: (1) NEW rows — each must include a one-line quote from the notes showing where the commitment appeared, (2) STATUS changes for existing rows if the notes prove DONE or STUCK, (3) MERGE suggestions if two rows are the same intent. If you cannot quote it, do not add it. End with OPEN_ROWS as a markdown table: Commitment | Owner | Source | Status | Next human step (one short sentence each).
Why it’s non-obvious: Without the quote rule, models happily invent plausible tasks that nobody actually agreed to. The quote requirement keeps the ledger evidence-based enough that finance or legal peers will not treat it as fiction.
2. One sendable follow-up per OPEN row (you send it)
OPEN_ROWS: [paste table]. For each OPEN row, draft exactly one Slack or email paragraph — calm tense, names the owner, references the source meeting date, proposes one concrete ask or deadline — ends with "Reply if this is already done." Do not send anything; I will paste manually.
Why it’s non-obvious: Autonomous nudging breaks trust fast. Auxot drafts; you send. Same human-in-the-loop bar as the rest of the tutorial library. The win is speed and wording, not surprise pings.
3. Priority drift when OKRs sit beside minutes
Add your quarterly priorities file (or bullet list) as another attached context, then:
Last four weeks of minutes are in the Minutes inbox file. Stated priorities are in the objectives context file. List (1) three commitments that align with priorities and moved forward, (2) three OPEN ledger rows that never appear in recent minutes despite being priority-aligned — probable drift, (3) one calendar or meeting-design change you recommend for next week. No motivational language — operational only.
Why it’s non-obvious: Most teams run either standup hygiene or strategy docs; rarely both in the same prompt. Letting the agent cross-read priorities and minutes side by side surfaces quiet deprioritization without a lecture.
Go deeper
Ledger vs rolling minutes
Keep raw minutes ugly if you need to. The ledger is the curated layer: one row per commitment, explicit owner, explicit status. The Meeting follow-through agent reconciles minutes → ledger, not the other way around. If you collapse them into one file, diffs get noisy fast.
Troubleshooting
- Too many OPEN rows every week: Your charter is probably too loose; tighten quote required, cap new rows per reconciliation, or split personal task noise into a second agent.
- Owners always read “TBD”: The model is guessing. Update the ledger manually after meetings so owners are explicit before you automate.
- People ignore the drafts: Shorten to one sentence plus an ask; long guilt paragraphs get skimmed.
Optional packaging
If the same rubric should apply across agents, capture it as a Skill (Create a Skill) so your Meeting Synthesizer and Meeting follow-through agents share one how we write commitments standard.
Walkthrough
Step 1: Create the ledger context file
Go to Settings → Context Files, add Commitments ledger (or your naming convention). Start from this skeleton:
| Commitment | Owner | Source | Status | Notes |
|---|---|---|---|---|
| Example: Ship pricing FAQ | Dana | Exec sync 2026-04-12 | OPEN | Waiting on Legal |
Replace the example with your first real rows, or leave only the header row until you have evidence.
Step 2: Charter the Meeting follow-through agent
In Chat with the Admin Agent, describe the agent in your own words, or paste:
Create a "Meeting follow-through" agent. It reads my Commitments ledger context file plus new meeting notes I paste. It never invents commitments: every NEW row must cite a verbatim short quote from the notes. It proposes ledger updates as markdown tables I can copy back into the ledger file. It uses statuses OPEN, STUCK, DONE only. If ownership is ambiguous, it flags AMBIGUOUS instead of guessing a person.
Answer follow-up questions until the agent appears under Settings → Agents.
Step 3: Attach context and run the first reconciliation
Attach Commitments ledger (and your minutes file if you maintain one). Paste fresh notes at the bottom of the chat. Run the delta prompt from The agent can do that? → 1. Review each NEW row: if you would not defend it in standup, delete it before copying into the ledger.
Step 4: Close the loop yourself
Pick the top one to three OPEN rows. Use prompt 2. for drafts. Then you send. Mark DONE in the ledger only after real-world follow-through actually happened (not when the draft looked polite).
Step 5: Add rhythm (optional automation)
If the same reconciliation should fire on a schedule, add Scheduled Tasks on the agent (Run an agent on a schedule) with a prompt that points the model at the latest minutes file contents; still no auto-send to Slack or email unless you deliberately wire a workflow (Run a workflow) with human approval (Require human approval before risky actions).
What’s next
- → Turn meeting notes into action items and decisions. Run this when a single meeting still needs a clean decisions-and-actions pass before it hits the ledger.
- → Keep your context files honest and fresh. When the ledger grows stale, these are the same refresh habits that keep agents trustworthy.
- → Set up your Monday morning briefing. Optional: fold OPEN row count and top drift into a weekly digest you already read.
- → Run a five-minute operator health check. Confirm System Health is trustworthy before you chase stalled commitments across meetings.
Reference
- Pages in Auxot: Chat, Settings → Context Files, Settings → Agents, Scheduled Tasks (optional)
- See also: Run the tools worker (auxot-tools), Run a five-minute operator health check, Add your first context file, Create a Skill, Trigger a workflow with an intake webhook (when minutes arrive from another system)