meeting-notes

---
name: meeting-notes
description: Turns a transcript or rough notes into a decision log and an action table where every action has one named owner, one deliverable and one date. Use after any meeting or call where things were agreed and nobody wrote them down.
---

# Notes and actions that survive the meeting

Two blocks: what was decided and why, and who does what by when.

## Claude does

1. Ask for four inputs: the transcript or notes, the meeting title and date, the attendee list with roles, and prior open actions. Refuse to start without the attendee list, because owners cannot be named from initials.
2. Read the whole input first. Mark every passage as decision, action, open question or background. Background is discarded.
3. Build the **decision log**: decision, rationale given in the room, dissent recorded, open issue left. If no rationale was spoken, write "rationale not stated" rather than invent one.
4. Build the **action table**: one named person, one deliverable, one date. Never a team, never "someone should", never "TBC". A row missing any of the three goes under "needs an owner", not into the table.
5. Apply the two-minute threshold. Under two minutes of the owner's time, draft it inline now; longer, it becomes a dated action.
6. Stay exception-driven. What changed, what is blocked, what is decided. Do not retell the discussion.
7. Carry prior open actions into a "still open" section with their original dates, so slippage is visible.

## Then Codex does

Codex takes the bulk pass: splitting the transcript into speaker-attributed segments, tagging each with one label from the closed list, pulling candidate action rows with their verbatim source line, and flagging rows missing an owner, deliverable or date. Repetitive work over hundreds of turns. Claude fills in {TRANSCRIPT_PATH}, {ATTENDEE_LIST}, {MEETING_TITLE_AND_DATE} and {PRIOR_OPEN_ACTIONS} in CODEX.md first.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. Every action row carries all three fields. Any row missing one goes to "needs an owner".
2. Every owner name appears in the attendee list. An invented owner is an automatic rerun.
3. Every decision and action cites a real source line. Spot-check three at random.
4. No speculation words: "probably", "presumably", "likely meant". The Amazon WBR sources prohibit speculation; it is a rerun here.
5. Dates are real calendar dates, not "next week".
6. Reject and rerun on any failure. Never hand-patch: a patched row hides which pass produced the error.

## Rules

- One item, one label, from the closed list. Nothing exits unlabelled.
- An action is one named person plus one deliverable plus one date, or it is not an action.
- Exception-driven only. No transcript retelling.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude does the segmentation and extraction pass itself and says plainly that it did both halves.

## Built from

- Anthropic, Skill authoring best practices, https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices - the normative SKILL.md spec.
- ComposioHQ/awesome-claude-skills, https://github.com/ComposioHQ/awesome-claude-skills - 74,501 stars via api.github.com; catalogues meeting-insights-analyzer.
- w95/awesome-claude-corporate-skills, https://github.com/w95/awesome-claude-corporate-skills - 189 stars via api.github.com; named-owner and DRI structure.
- Commoncog, The Amazon Weekly Business Review, https://commoncog.com/the-amazon-weekly-business-review/ - exception-driven, no-speculation discipline.
- Superhuman, The Inbox Zero Method, https://blog.superhuman.com/inbox-zero-method/ - forced-choice taxonomy, two-minute threshold.

# Notes and actions that survive the meeting - Codex task

Bulk extraction pass over a meeting transcript. You have no memory of any prior conversation. Everything you need is below.

## You are given

- `{TRANSCRIPT_PATH}` - the raw transcript or notes file in the working folder.
- `{ATTENDEE_LIST}` - the only names that may appear as an owner, each with a role. Nobody outside this list exists.
- `{MEETING_TITLE_AND_DATE}` - used to resolve relative dates such as "by Friday".
- `{PRIOR_OPEN_ACTIONS}` - carried-over actions with owner, deliverable and original due date. May be empty.

## Produce

Three files in the working folder.

1. `segments.md` - the whole transcript split into speaker-attributed segments in original order. One line each: `[n] SPEAKER | LABEL | verbatim text`. LABEL is exactly one of `DECISION`, `ACTION`, `OPEN_QUESTION`, `BACKGROUND`. Every segment carries a label. Nothing is dropped.

2. `candidate-actions.md` - columns: Owner, Deliverable, Due date, Source segment, Missing fields. One row per `ACTION` segment. A field you cannot read from the transcript stays empty and is named in Missing fields. Do not guess. Append `{PRIOR_OPEN_ACTIONS}` unchanged under a `## Still open` heading, keeping the original dates.

3. `candidate-decisions.md` - columns: Decision, Rationale, Dissent, Open issue left, Source segment. One row per `DECISION` segment. Where no rationale was spoken, write exactly `rationale not stated`.

## Rules

- Every row cites a source segment number that exists in `segments.md`. No citation, no row.
- Owner names must match `{ATTENDEE_LIST}` exactly. Initials, first names alone or job titles go in Missing fields instead.
- A due date is a real calendar date resolved against `{MEETING_TITLE_AND_DATE}`. "Next week" and "soon" are not dates: leave empty and flag.
- A team, a department or "someone" is never an owner. That is a missing field.
- Never invent a fact, a number or a quote. Verbatim text is copied, not paraphrased.
- No speculation words anywhere: probably, presumably, likely, seems, appears to.
- Public information only. Look nothing up outside the supplied files.
- If the transcript is empty or unreadable, stop and say so rather than producing empty tables.

## Return

Print to stdout, in order:

1. The three filenames written, each with its line count.
2. Total segments, plus counts for DECISION, ACTION, OPEN_QUESTION and BACKGROUND. The four must add up to the total.
3. Candidate action rows, and how many have at least one missing field.
4. Any owner name found in the transcript that is not in `{ATTENDEE_LIST}`, listed verbatim.
5. The line `SPECULATION WORDS: 0`, or the offending lines if any slipped through.