Step Six Is A Research Preview, And The Command On The Card Will Not Run
This card is better than most. Six steps, each with a real prompt you could paste, building toward something concrete: your exported chat history turned into an Obsidian vault, wikilinked, with a wiki layered on top and your apps feeding it. Steps one to three are the kind of thing you could do this afternoon and still have working in two years.
Then step six says: “Run claude --channels to connect iMessage. Text your own number and Claude Code replies from your wiki.”
That flag is real. It is also a research preview, it does not appear in claude --help, and it takes a required argument the card does not print. Run it exactly as written and nothing connects.
Source card: “Build your Claude second brain” — BANI Academy, 18 October 2026. Six steps: export and organise, open Graph View, build the wiki, add MCP connectors, set up slash commands, text your brain.
View the original post on Facebook
The command as printed is missing its argument
Here is the flag in the Claude Code CLI reference, quoted in full:
“
--channels— (Research preview) MCP servers whose channel notifications Claude should listen for in this session. Space-separated list ofplugin:<name>@<marketplace>entries. Requires Anthropic authentication through claude.ai or a Console API key.”
So the flag names which channel to listen to. Bare, it names nothing. The working command for the thing the card is describing is:
claude --channels plugin:imessage@claude-plugins-officialAnd that is the last line of a five-step setup, not the first. The documentation lists what has to be true before it works.
| What the card says | What the documentation says you need first | What happens without it |
|---|---|---|
“Run claude --channels” |
A macOS machine. “The iMessage channel reads your Messages database directly and sends replies through AppleScript. It requires macOS” | Nothing to run. Windows and Linux are out for this channel |
| — | Bun installed. “Each supported channel is a plugin that requires Bun” | The plugin cannot start |
| — | The plugin installed: /plugin install imessage@claude-plugins-official, possibly after adding the marketplace |
--channels accepts the name but the channel never registers |
| — | Full Disk Access to ~/Library/Messages/chat.db |
Documented plainly: “Without this, the server exits immediately with authorization denied” |
| — | A macOS Automation approval the first time Claude replies, letting your terminal control Messages | The inbound message arrives; the reply does not go out |
| “…to connect iMessage” | claude --channels plugin:imessage@claude-plugins-official — and on Team or Enterprise, an Owner must have switched channelsEnabled on |
Claude Code starts normally and the channel silently does not register |
“Text your brain” means your Mac has to be awake
The mock-up on the card shows a phone: “My Brain · Active”, with a reply that reads your Granola notes and Gmail. It reads like a service. It is not one. From the same documentation page:
“Events only arrive while the session is open, so for an always-on setup you run Claude in a background process or persistent terminal.”
There is no server in this picture. Your laptop is the server. Text your number while the lid is shut and the message sits in Messages until you next start a session — and because the plugin reads the database rather than subscribing to a stream, what you get on reconnect is whatever it finds, not a replay of a queue.
This is worth being clear about because it is the difference between a feature and a habit. “Text your brain” works if you are the sort of person who leaves a terminal running on a desktop Mac. If your machine is a laptop that sleeps in a bag, steps five and six are a demo, not a workflow.
What each step actually asks for
The card’s fine print says “keep private data private”, which is the right instinct. It is worth laying out what the six steps add up to, because each one is reasonable on its own and the total is larger than any of them.
| Step | What it asks for | What that opens |
|---|---|---|
| 1. Export + Organise | Your full ChatGPT and Claude history, unzipped into a folder | Every conversation you have ever had, in plain text, on disk, unencrypted |
| 1. (the prompt) | “Launch 10 sub-agents to go through every conversation” | Every one of those conversations read end to end by a model, billed by the token. The card gives no estimate of time or cost |
| 4. MCP connectors | Gmail, Granola, NotebookLM, Notion, Slack, Calendar | Six more authorisations, and the contents of your mail, meetings and messages flowing into the same folder |
| 6. iMessage channel | Full Disk Access to ~/Library/Messages/chat.db |
Your entire Messages history — not just the thread with yourself — readable by the process |
| 6. (self-chat) | Nothing. “Texting yourself bypasses the gate automatically, with no setup” | Convenient, and it means the sender allowlist protects you from everyone except anyone holding a device signed into your Apple ID |
| 6. (permission relay) | Optional, if the channel declares it | Documented bluntly: “Anyone who can reply through the channel can approve or deny tool use in your session” |
To be fair to the card: its footer does say “Review what the AI writes, and keep private data private”, which is more than most cards of this kind say. What it does not say is which step is the expensive one.
Step one has a 24-hour fuse
The card prints two timings in step one: ChatGPT “≈ 3 days”, Claude “≈ 5 min”. Neither vendor publishes a figure. Anthropic’s help page says only that “there may be a small delay while the export is generated”. So treat the contrast as one person’s experience rather than a specification — it may well be what they saw, but it is not a number you can plan around.
The detail that does matter in step one is one the card omits. From the same help page: the export arrives as a download link by email, and the link expires 24 hours after delivery. You have to be signed in to use it, and if it lapses you request a new export and wait again. Also worth knowing before you start: “It’s not possible to run an export from Claude for iOS or Android.”
Which means the practical order is: request the export when you are at the machine you intend to build on, and have the folder ready. Not before.
Diagram: which steps survive if channels change tomorrow
Research preview is a specific status, and the documentation spells out what it implies: “Availability is rolling out gradually, and the --channels flag syntax and protocol contract may change based on feedback.” So the honest way to read this card is to ask which of the six steps still stands if that happens.
| Step | What it depends on | If the preview changes or goes away |
|---|---|---|
| 1. Export + Organise | A zip file and a folder of markdown | Unaffected. Plain files on your disk |
| 2. Graph View | Obsidian, a local app | Unaffected. Obsidian is not an AI product |
| 3. The wiki | Claude Code writing markdown into folders | Unaffected. Worst case you write the pages another way; the pages remain |
| 4. MCP connectors | MCP, which is stable, plus six third-party services | Mostly fine. Any one connector can break without taking the vault with it |
| 5. Slash commands | Claude Code commands — stable. But “run each from your phone via iMessage” is step 6 | The commands survive; the phone part does not. You run them at the keyboard |
| 6. Text your brain | The research preview, entirely | This is the step that goes. Nothing else on the card breaks with it |
That ordering is the actual advice. Steps one to three are the second brain. They are a folder of markdown, a graph view over it, and a wiki you can read without any AI running at all. Steps four to six are conveniences layered on top, and the last one is the only part flying on a preview.
It also answers the caption’s premise. “Your AI forgets every chat the second you close it” is doing a lot of work — the same page published a card a fortnight earlier listing Memory as a four-star, two-minute settings toggle that makes Claude “recall every past chat so you stop re-briefing from scratch”, which we went through in the piece on that card. The honest case for building a vault is not that the AI forgets. It is that a folder of markdown is yours, and a chat history on someone’s server is not.
Four honest caveats
- We did not build it. We have no Mac running this channel. Everything above is the card checked against Anthropic’s published documentation, not a report from having tried it.
- A research preview moves. The channels page says the flag syntax may change. It is entirely possible that by the time you read this the command is shorter, or that
--channelshas graduated. Check the page before concluding the card is wrong rather than early. - The export timings may be real experience. We say the vendors do not publish those numbers, not that the card invented them. One person’s export genuinely can take minutes or days depending on account size and queue.
- “Ten sub-agents over every conversation” has a cost we did not compute, because it depends entirely on how much history you have. We flag that the card gives no estimate; we are not supplying one, and if you have years of chats you should run it on a subfolder first.
The one-line version
Steps one to three build something that is yours and keeps working; step six rests on a research-preview flag that is macOS-only, needs Full Disk Access to your entire Messages database, only delivers while a terminal is open, and takes an argument the card does not print.
Build the vault. Add the connectors. Treat the phone part as the experiment it is labelled as.
Source card: “Build your Claude second brain” — BANI Academy, 18 October 2026.
View the original post on Facebook · Related: checking the version number on a card before you follow it
