Get started
What it plugs into
The component pages describe roles, because that is what keeps the design portable. This page answers the other question: what do I actually install, and where does the thing I already use fit?
- Required The runtime Where the skills are loaded and run Claude Code · Claude Cowork
- Swappable The ingestors How the text gets in. Pasting is enough paste a VTT · Deep Thought · Klang · Fireflies · a local model
- Required Your vault The folder your material lives in any folder · Obsidian to read it in
- Swappable The mirrors Systems that get a copy, never the truth Todoist · GitHub · Jira · chat
The runtime — required
Everything here runs inside a Claude session. A skill cannot start one; it is loaded into one that is already running. There are two interfaces, and they run the same skills against the same files.
| Claude Code | Claude Cowork | |
|---|---|---|
| Where | a terminal | Claude Desktop, macOS and Windows |
| You type | /ops <transcript> | what you want, in plain language |
| Install | a symlink per skill | the same folder, no symlinks to think about |
| Suits | anyone comfortable in a shell | anyone who is not |
| Needs | a Claude plan | a Pro, Max, Team or Enterprise plan |
Get Claude Code · its documentation
The choice is not architectural. Nothing in this manual is different depending on which one you use — the pages show the command because a command is unambiguous, and in Cowork you describe the same thing instead of typing it.
The ingestors — optional, and pasting is one
The simplest ingestor is your clipboard. Export the transcript your
meeting tool already produces — a .vtt from Teams, Zoom, Meet, anything —
drop the file in the inbox or paste it into a session, and the framework processes it.
No integration, no API key, no account anywhere.
That is worth stating plainly because the rest of this section can read like a list of requirements. It is not. These are only the ingestors: they get text in, and nothing else. Every one of them is a convenience over pasting.
- MCP — the open protocol a session uses to reach a provider directly, so you skip the export-and-paste step.
- Deep Thought — speech to text with speakers and search, over MCP.
- Klang — a Swedish provider doing the same job, also over MCP.
- Fireflies — publishes a documented MCP server, so a session can fetch a meeting without leaving the session.
- A local model — Whisper or similar. No network, no account, worse transcripts, identical processing afterwards.
- Anything else that exports text. Otter and most meeting tools produce a transcript file; that file is a valid ingest. (Otter may well offer an MCP server too — I could confirm one for Fireflies and not for Otter, so this page does not claim it.)
Whichever you use, the distinction on the two-hubs page is the one to read: the ingestor produces text, and the framework decides what the text means. Only one of those is replaceable, and it is not the framework.
The vault — required, but it is just a folder
A folder of Markdown files on your own machine. No database, no service, no account. The conventions — which folders exist and what each means — are declared in the contract.
Obsidian is a pleasant way to read and link that folder, and the reason machine-written folders start with a dot is partly that Obsidian hides them. It is not required. Any editor opens the same files.
The mirrors — swappable, and never the truth
Several systems get a copy of something the vault holds. The direction matters more than the product: the vault is the source, the external system is a mirror, and when they disagree the vault wins.
- Todoist — a task ledger reachable from a phone. One script mirrors outward; see the task pipeline.
- GitHub and Jira — metadata only, never the code and never an issue's comment thread. See the repository archiver and the board archiver.
- A chat platform — read and send, through a client that knows nothing about vaults. See the messaging client.
What you can download, and what you cannot
| Part | Where |
|---|---|
| The framework itself | tandregbg/core-claude-skills — MIT |
| The dashboard, capture app, archivers, task pipeline | not published. The pages describe them because the manual covers the whole system, and the pattern transfers even when the code does not |
| A March snapshot of the dashboard | public, and six months stale. The current version is not that snapshot |
Where the thinking comes from
The arguments this manual states as mechanism are argued at length, in first person, elsewhere. Tomas André's insights are where a claim like a transcript is an anecdote, a diarised series is data gets made rather than asserted.
And if you would rather be taught this than read it: Mindtastic runs workshops on working this way. That is a separate thing you pay for; everything on this site is free and stays free.
From ecosystem.yaml, contract 41 · core-skills 1.89.4 · read at build 2026-10-08