claude-transcript-skill

Use this when you finish a meeting: organize everything into what the AI actually understands and what it does not, checked against your own files, repositories, docs, and designs.

A transcript is not a spec. This skill turns a meeting into a compact decision list, separates what is already true in your codebase from what is only a proposal, and keeps asking until nothing meaningful is left ambiguous.

What it does

• Captures every distinct point, including side comments, corrections, and admin items.
• Verifies each point against the real repo, branch, canonical docs, and design artifacts.
• Marks each item as implemented, documented only, design only, proposed, or still unverified.
• Reports only what advances a decision, in plain language, with evidence and caveats.
• Narrows with you over a few short rounds instead of repeating the whole audit.

Install

Copy this repository into your skills directory, or clone it directly.

Claude Code:

git clone https://github.com/Mpayro/claude-transcript-skill.git ~/.claude/skills/transcript-exploration

Codex:

git clone https://github.com/Mpayro/claude-transcript-skill.git ~/.codex/skills/transcript-exploration

Any other assistant that reads a SKILL.md: clone it anywhere and point the assistant at the file.

Subagent contract

The skill delegates verification to subagents. By default it spawns subagents of the same AI that is running it, one tier down: Claude spawns Claude subagents, Codex spawns Codex subagents, and any other assistant spawns subagents of itself. The cheapest model that can do the lookup is chosen, and the model is passed explicitly on every spawn.

The user's explicit choice always wins. Name the provider and model and the skill uses exactly that.

Files

• SKILL.md - the skill itself.
• agents/openai.yaml - Codex display name and default prompt.