The AI Meeting Minutes Writer skill for Claude

You probably recognise the problem the skill solves in two forms. The first is time: after every meeting someone spends an hour typing up minutes that nobody then reads. The second is worse and more subtle: the minutes exist, but they are a chronological retelling. Whoever reads them still does not know, after ten paragraphs, what was actually decided and who is doing what. Two weeks later the same discussion starts again, because nobody can point to where the last one ended.

Both of those are tackled here: writing them up costs you a prompt instead of an hour, and the result is not a retelling but a decision document with a fixed place for decisions, action points and open items.

It is written for anyone who chairs or writes up meetings: business owners and their teams, project managers, office managers, committee boards and freelancers who record client conversations. Together with the rest of the skill library NL, we make it available for free from our AI and automation service, with no account and no sales email afterwards.

One thing you should know beforehand: the skill takes the minutes, it does not chair the meeting. The quality of the minutes stands or falls with the quality of the source. A meeting where nobody decides anything produces minutes with an empty list of decisions, and that is exactly what ends up in the write-up. It never gets made to sound better than it was, and that is precisely why you can build on it.

The write-up is a retelling. Then Pete said this, after which Karen remarked that, and a discussion followed. Chronology feels thorough, but the reader is not after a film, they are after the outcome. That is why every write-up from this skill opens with a summary of five lines at most: what has been decided, what needs to happen, what is still open. Whoever reads only those five lines knows enough. That is Minto's Pyramid Principle, applied to meetings: conclusion first.

Proposals quietly turn into decisions. The most dangerous sentence in a meeting is "we were all agreed on that, weren't we". A decision that was never made then appears in the minutes, and three weeks later someone feels blindsided. Hence the explicitness test: a decision only goes into the list of decisions if it was explicitly made in the source. Someone proposes it and there is agreement, or the chair confirms it. Everything short of that is an open item, with the reason next to it.

Action points without an owner and without a moment. "I will take a look" is the kiss of death for any action point. Getting Things Done by David Allen has been saying it since 2001: an action without an owner is a wish, an action without a next step stays put. That is why the skill forces an owner and a deadline for every action point, and where the source does not contain those, it does not guess but writes [FILL IN], with a note on which fragment was unclear.

Quote and interpretation run into each other. There is a real difference between "Sanne said the newsletter costs more than it brings in" and "the newsletter costs more than it brings in". The first sentence is a record, the second is a conclusion nobody has drawn. The two stay strictly separate: literal statements sit between quotation marks with a name attached, and anything the skill infers or summarises is marked as interpretation.

The write-up passes judgement. "Heated discussion", "not a very constructive attitude", "well prepared": those are judgements, not facts. They belong in a performance review, not in minutes the whole team reads. What was said and decided goes in, and the interpretation stays with the people themselves.

In diff form, with lines from the skill's own example: what goes out and what comes back in its place.

Then comes the core: a fixed minute-taking logic of six steps. First the skill reads the whole source and establishes who is speaking; for a transcript without speaker labels it asks whether you can name them, and if not, those statements are recorded as speaker unknown rather than attributed to the wrong person. Then it runs through an input checklist and asks at most three targeted questions about what is missing; the rest it covers with [FILL IN].

Step three harvests decisions and action points with the explicitness test: if in doubt, it is an open item. Step four separates quote from interpretation. Step five builds the write-up in the eight fixed blocks. And step six is a mandatory self-check pass: does every action point have an owner and a deadline or a [FILL IN], is every decision traceable to a moment in the source, is there no statement in it that is not in the source, and has every judgement about people been removed.

The input checklist itself is short: the source (mandatory, because without a source the skill stops), the name and date of the meeting, who was present and absent, whether any passages are confidential, who the minutes are for, and whether there is a previous list of decisions or actions that needs updating. That last one is quieter than it sounds but worth its weight in practice: the numbering (B1, A1) exists precisely so a following meeting can refer back to it.

The output format sits in the file as a literal template, with the rules that go with it: the summary is never longer than five lines, quotes sit between quotation marks with a name attached, interpretation gets marked, and an empty block stays in place with "none". Then comes the refusals block with seven lines, which are covered one by one below in Section 04. The file closes with the shortened example you saw in FIG.01 and a list of five sources, from Minto to Rogelberg. More on that below.

The summary at the top comes from Barbara Minto. Her Pyramid Principle, developed at McKinsey and described in 1987, says: start with the conclusion, not with the build-up. For minutes that means: whoever reads only the first five lines has to know what was decided and what needs to happen. Want to apply that way of thinking more broadly than write-ups, there is a separate Pyramid Principle Writer skill.

The action points come from David Allen and George Doran. Getting Things Done (2001) provides the rule that every action needs an owner and a concrete next step: an action without an owner is a wish. And the SMART criteria, described by Doran in Management Review in 1981, provide the requirement that an action point be worded in a checkable way. Anyone who wants to go deeper into that can turn to the Getting Things Done Coach skill and the SMART Goals Coach skill.

The why comes from Rogelberg and Lencioni. Steven Rogelberg (The Surprising Science of Meetings, 2019) and Patrick Lencioni (Death by Meeting, 2004) both describe the same mechanism from a different angle: a meeting without recorded outcomes and without follow-up loses its value, however good the conversation was. The minutes are not a by-product of the meeting; they are the proof that the meeting led somewhere.

The separation between quote and interpretation is the oldest convention of the craft itself. A record documents, it does not judge. That rule does not sit in any one book because it is older than the books; every clerk and every court reporter works by it. Here it has simply been made explicit and checkable, because an AI would otherwise breach that rule just as easily as a rushed human minute-taker.

Even without installing the skill, you can use these principles to improve your own write-ups: put the conclusion at the top, give every action point an owner and a moment, and never write down a decision that was not explicitly made.

After that, using it is simple: paste the transcript or your notes and ask for minutes. It asks about anything missing itself, with a maximum of three questions. If you are new to skills, first read what Claude skills actually are NL and how to install a skill NL; if you want to build skills yourself afterwards, the method is in writing a SKILL.md yourself NL. The broader explanation of working with AI is in the knowledge base.

There are also situations where you are better off leaving it be. Minutes with statutory or constitutional requirements are one of those: the minutes of a shareholders' meeting, a Dutch VvE (owners' association) meeting or a works council follow a set of rules that prescribes the form, and that should be checked by a person, with a lawyer where needed. The file says exactly that in so many words. The second limit is the source: without a transcript or notes there is nothing to write minutes from, and it does not make things up. Reconstructing minutes afterwards from your memory is possible, but then they are your notes and the skill treats them as such.

And a limit that sits outside the skill but that we are honest about: watch what you send to an AI service. A transcript of a meeting about people, clients or confidential figures quickly falls under your organisation's AI or privacy policy. Keeping passages out of the write-up is possible on request, but whatever you paste in is already shared. If in doubt: check the policy first, then paste.

Finally: what gets minuted is the meeting you had, not the meeting you wished you had had. If your meeting produces no decisions, the write-up gets an empty list of decisions. That feels confronting, but it is also exactly the signal you can act on: maybe the write-up is not the problem, the meeting is. For client conversations and intakes there is, incidentally, a separate skill that works on the same principle, the Intake Summary skill: same muscle, different conversation.

If more runs through your organisation than just meetings, think of figures, project progress and internal questions that keep getting asked again, that is exactly the work of the Reporting employee: it prepares weekly and monthly reports and a person gives approval before anything goes out the door. Read what an AI employee is and does, or first teach your team to work with this kind of skill themselves through the AI training NL.