Daniel Cohen

Career

Meeting review

Get one review of how you work in meetings, from three months of your own transcripts and backed by your own dated quotes.

Why it worksFew people read back the transcripts that record how they decide and follow up.
  • It counts decisions made versus deferred and follow-ups closed.
  • Every finding quotes your own words, with a date.
  • It ends with three habits to test next month, each with a count to check.
  • It never scores anyone else.
15 minClaude CodeClaude appOpen meeting-review/ on GitHub
meeting-review-2025-06-27.mdSample, fictional
Alex, Apr '25 to Jun '2534 meetings read

What to change: you defer without a date

12 of 31 decision points were deferred. 7 of the 12 had no date.

May '25, team sync: "Let's pick this up next week."
May '25, planning: "I'd like to think about pricing a bit more before we decide."
Jun '25, team sync: "Park it, we'll circle back."

Of the 7 undated deferrals, 3 came back in a later meeting and 4 left no trace in these transcripts.

Try next month

Replace "park it" with a date and a bringer. Baseline 7 of 12 undated, aim for 0.

Fictional sample. Alex, the team and every quote are invented. The full fictional review is in the repo.

What you get

One review of about 3,000 words, built from your transcripts and read in date order.

  • What you do well and what to change, each backed by 2 or 3 dated quotes in your own words.
  • PM craft and team-lead craft: problem framing, use of evidence, decisions made versus deferred and follow-ups closed, then feedback, delegation, unblocking, airtime and how you run the meeting.
  • Your first month against your last, with counts.
  • Three things to try next month, each with a count to check it against.
  • The skill marks anything it infers as inferred.
  • Built as a general method: test it on your own transcripts. Work done outside meetings needs a separate check.

Copy it

Time to set up: 15 minutes, most of it the export. Pick the assistant you use. The choice is remembered on this device.

Setup steps

  1. Export your transcripts. Most meeting tools export .vtt, .txt or .docx, and a .md file works too. The speaker names must be in the file, because the skill has to tell your words from the others'. Three months gives a good read. Fewer than 6 meetings is too thin for patterns over time.
  2. Put them in one local folder, for example ~/meeting-transcripts/. A date in each file name, such as 2025-04-17-team-sync.vtt, saves the skill from guessing.
  3. Copy the skill with the command in the Claude Code or Claude app tab above.
  4. Run it. It asks which speaker labels are you and whether you want the PM view, the team-lead view or both. With more than 10 meetings it starts fast-model subagents to read the batches, and the main model writes the review. Before it writes, it checks every quote against its source file with an exact-text search, drops any that do not match and reports how many it dropped. It saves the review next to the folder as meeting-review-<date>.md.
  5. Clean up. Delete the transcript folder, which also holds _review-notes/. The review stays.

Try a review

First prompt
Use meeting-review on the transcripts I supplied. Ask which speaker labels are mine and what date range to assess. Use the PM view. Check every quote against its source and state where the evidence is too thin.

Expected result: a review with dated quotes, the number of meetings assessed and specific habits to try; check a cited quote against its transcript.

Where it works

AssistantSupport
Claude CodeThe full version. It reads the folder, runs subagents for the batches and saves the review as a file.
Claude app (desktop and web)Upload the zip as a skill and attach your transcripts. There are no subagents, and the number of transcripts that fit in a chat depends on their length.
Other assistantsThe skill is written for Claude. Pasting it into the custom instructions works for a small set of transcripts. The folder read, the subagents and the saved file are Claude Code features.

Privacy

Transcripts hold other people's words and personal data. Read these before you export anything.

  • Company policy and consent. Check your company's policy on recording, on storing transcripts and on using AI tools. Check the consent rules for the people in your meetings. Rules differ by country and by company.
  • Keep the export local. Use a folder that no shared drive or sync service picks up. The skill reads only the folder you name and searches nothing. The text still reaches the model provider you use with Claude, so the policy check covers that too.
  • Delete the export. Remove the transcript folder and _review-notes/ when you have the review.
  • The review describes you only. It quotes only your words, names other people by role, and skips candidate interviews and pay or performance conversations unless you say otherwise.
  • Read it before you share it. The review quotes your meetings and may carry company details.