Back to storiesPractical guides

A meeting-notes agent: turn discussion into follow-up without inventing assignments

Work through two public meeting records, split a four-step procedure, preserve unknown owners and dates, and produce a reviewable register plus an unsent follow-up. Download the actual outputs.

8 min readMarkdown
On this page7 sections

A suggestion is not an assignment

A meeting note says a team could run a screen-sharing session next time, if there is time. A polished summary can turn that into a task with a named owner and a deadline. Neither detail is necessarily in the meeting.

We used two public Python documentation community records to build a follow-up register. The reviewed output contains 11 selected records: two reports and nine items that need confirmation. There are no explicit assignment candidates. That empty assignment list is useful: it shows what the meeting notes do and do not authorize.

Use real minutes and keep their date attached

The inputs are the September 1, 2026 minutes and October 6, 2026 minutes. We read both in full and selected passages about potential follow-up, a migration procedure and reported progress. This is an archival exercise, not a complete summary or a statement about the community’s current tasks.

The extraction, editing and file export happened in this Codex conversation, using Python for the export. We did not run Claude Code or Onevium for this exercise, attend the meetings, contact participants or create tasks in their systems.

Download the working pack. It includes bilingual selected-input paraphrases, a reusable prompt, first and reviewed records, a second-input extraction, CSV output and unsent follow-up drafts. Source URLs are pinned to a repository commit; the source repository uses CC0-1.0. Names and accounts are omitted from the exercise inputs.

Ask for evidence before asking for a to-do list

The prompt asks the assistant to classify each item before describing the next step. A report records something said to have happened. A proposal records an idea. A procedure describes how work could proceed. An assignment needs evidence that someone actually accepted or was given the work.

Use fields for source passage, classification, action or report, owner evidence, due date, dependencies and a confirmation question. Leave unsupported fields empty. The person speaking is not automatically the person responsible; a meeting date is not a delivery date.

Read these minutes and return a follow-up register.
Separate reports, requests, proposals, procedures and explicit assignments.
For every item, include a source reference and a confirmation question.
Only fill owner and due date when the source explicitly supports them.
Preserve conditions such as “if time permits”.
Split multi-step procedures without treating them as authorization.
Draft the follow-up, but do not send it or create tasks.

For a real team, add its task definitions and who may approve assignments. Missing dates can remain open if the team permits that; inventing a date never resolves the uncertainty.

The first edit: make the migration sequence reviewable

The September discussion describes updating redirect rules and obtaining approval, merging documentation changes and backports, merging redirects, then clearing the CDN cache after checking with a contact.

Our first register bundled that procedure into one entry. We revised it into four entries so each stage can have a separate confirmation question and, eventually, an owner. The dependencies retain the order described in the minutes. They are not an independently tested deployment runbook.

EntryWhat needs confirming
Prepare rules and obtain approvalWho prepares them, and what evidence confirms approval?
Merge docs and backportsWho checks the previous stage before merging?
Merge redirectsWho confirms the documentation stage is ready?
Check and clear cacheWho confirms readiness and carries out the cache step?

That edit changes the September register from five entries to eight. It adds neither owners nor deadlines and does not claim the underlying work was performed. The first file and the review note remain in the pack.

Actual review register: four procedure steps have no specified owner or due date
Reformatted from the exported register. Eleven records do not mean eleven assigned tasks.

A second input tests whether the distinctions still help

We applied the same approach to three selected passages in the October record. The translation update reports that the work met core requirements and was waiting for a core developer to move official inclusion forward. We kept that as reported waiting, not completed inclusion. A possible shared entry point to the developer guide remains a proposal.

The notes also describe an issue/PR-review walkthrough. We retained it as a report, rather than assigning the walkthrough again. It does not prove that every conditional proposal in September was completed.

This second input was processed in the same conversation. It is a second worked example, not a blind evaluation of model accuracy. The reusable part is the distinction and evidence field; the reader still needs to check the interpretation against the source.

The deliverable is a register and a short confirmation draft

The combined CSV has 11 records, nine requiring confirmation and two reports. All due dates are unspecified. The assignment-candidates file is empty because none of these selected passages supplies an explicit assignment with an owner. The follow-up draft asks who accepted the translation handoff, whether the common-entry-point proposal is approved, and who owns each migration stage.

One excerpt from the draft:

Please confirm separate owners and approval evidence for preparing rules, merging docs/backports, merging redirects and clearing cache. The proposed triage session remains conditional on available time and an agreed facilitator.

This draft was not sent. In a current project, review it with the meeting organizer and check present status before acting on historical notes. A short set of answerable questions is often more useful than a long list of confident but unsupported tasks.

Reproduce the output, then try your own meeting

Unzip the pack and run this command inside its meeting-followup folder, using Python 3.10 or later and a new output directory:

python3 export.py september-reviewed.json october-extraction.json --out my-output

The script checks duplicate record IDs, source references, required owner evidence and dependency cycles, then writes register.csv, summary.json and assignment-candidates.json. Compare them with expected/. It does not call an AI model, check current project status or prove that a classification is semantically correct. Repeating the export reproduces files; it does not repeat the extraction.

For another meeting, give the assistant only authorized notes and the prompt, then review its output against the notes. Start with three checks: did it convert a suggestion into a commitment, mistake a speaker for an owner, or fill an unknown date? Keep the unanswered questions visible until someone with authority resolves them.

The files and review guide explains the corresponding review surface in Onevium. The exercise itself remains a Codex-produced artifact.