Skriver

ai meeting note taker • meeting transcript tool • tool + skill

Skriver is an AI meeting note taker and transcriber with the evidence your agent actually needs.

Skriver records a meeting transcript, builds a clean meeting notes evidence folder, and teaches the agent how to turn that first pass into the best final transcript, summary, actions, and decisions without bluffing.

  • Transcribe audio and video into a conservative first-pass meeting transcript.
  • Capture the extra evidence a serious AI note taker should have: notes, screenshots, OCR, context, and diarization.
  • Notes are higher-trust clarification than OCR.
  • Add notes with --notes-file ./notes.md so they become part of the evidence bundle.
  • Works with Codex, Claude Code, and other agents that can read markdown repo instructions.
  • The agent’s second pass is where the best meeting notes and final summary are made.
*

The Tool

Make the best possible starting point for a meeting transcript.

Skriver runs locally on audio and video, writes a readable first-pass transcript, and gives the agent the files a real AI meeting note taker should have: notes, low-confidence segments, screenshots, OCR, context, and optional diarization.

run.json

The run contract. It tells the agent what completed and where to look next.

meeting-transcript.md

A conservative first-pass meeting transcript. Good enough to start from. Honest enough to trust.

evidence/

The real payload: meeting notes, OCR, screenshots, context, and diarization for the second pass. Notes added with --notes-file are especially important evidence.

//

The Skill

Learn how to turn the bundle into the best final notes.

This repo also ships a checked-in skill for agents: skills/skriver/SKILL.md. The repo also includes a focused review skill at skills/skriver-evidence-review/SKILL.md. Together they teach the review order, how to weigh notes against OCR, how to use screenshots and diarization as evidence, and how to create the best final transcript, meeting notes, and summary.

Why it matters

The best output does not come from the first transcript alone. It comes from the agent working the evidence folder carefully and turning it into better meeting notes.

How to load it

Codex can use the checked-in skill. Claude Code can use CLAUDE.md plus the repo skill in .claude/skills/. Other agents can follow AGENTS.md.

What you get

A better final transcript, stronger meeting notes, cleaner actions, and less hallucinated certainty.

>

Install

Install the tool globally.

macOS / Linux

curl -fsSL https://skriver.ansund.com/install.sh | bash

npm

npm install -g github:ansund/skriver

Windows

irm https://skriver.ansund.com/install.ps1 | iex

Load the workflow

skriver help agents

That prints the install path, usage path, and where the skill/workflow lives for Codex, Claude Code, and generic agents. The CLI is installed globally. The skill is checked into this repo and loaded from markdown.

>>

Use

Use Skriver like an AI note taker with receipts.

1. Check the machine

skriver doctor
skriver setup

2. Run on a file

skriver /absolute/path/to/meeting.mp4 --notes-file ./notes.md

Recommended: put your human meeting notes in a `.md` file and attach them as evidence.

3. Optional glossary

skriver /absolute/path/to/meeting.mp4 \
  --notes-file ./notes.md \
  --glossary ./team-glossary.txt

A glossary helps a meeting transcriber handle abbreviations, product names, and English terms inside Swedish or other languages more reliably.

4. Load the skill

Codex: use skills/skriver/SKILL.md
Claude Code: use .claude/skills/skriver/SKILL.md
Other agents: read AGENTS.md

The workflow is plain markdown on purpose, so it can travel across agent runtimes instead of being trapped inside one vendor feature.