Onboarding notes Claude Code writes for a new developer

Claude Code and Pando, for onboarding a new developer.

A new developer spends the first weeks asking where things are and why they are that way, and the answers land in a chat that scrolls away. Claude Code can read the repository and answer those questions, but it tells only whoever is running it, and what it learns stays on that machine: in Anthropic’s words, "Auto memory is machine-local."

Yes, and where it writes them decides who can use them: in CLAUDE.md they reach every Claude Code session through the repository, and in a branch of your Pando outline that you share, the new developer reads them in a browser and writes questions under the lines they are about, which Claude Code answers the next time you run it and ask.

Set it up

In a terminal: claude mcp add --transport http --scope user pando https://pando.ink/mcp
Start claude, run /mcp, select pando and choose Authenticate.

On Pando’s consent page choose Read and write your outline. Approved with Read your outline only, it can change nothing outside its own memory bullet, so hold it to the branch (below) before its first run. Either way it starts with your whole outline.

What it writes

Make a bullet for it, Onboarding: billing-service, then start Claude Code in the repository and ask it to write onboarding notes for a new developer under that bullet in Pando: where things live, each decision with its reason in the note, and for how to run it, a pointer to the README. A week later the branch reads like this:

- Onboarding: billing-service
  - Where things live
    - src/api: the HTTP routes
    - src/jobs: the nightly runs
  - How to run it
    - See the README
  - Decisions, and why
    - Postgres, not DynamoDB
      note: The reports join across accounts. Decided March 2026.
      - Why not a read replica for the reports?
        - The month-end report reads the nightly rows at once, and a replica can lag.

The question is the new developer’s, written under the decision it is about. The line under it is Claude Code’s, written the next time you asked it to answer the open questions. Pando does not tell Claude Code who wrote a line, so agree that a question ends in a question mark and has nothing under it yet, and say so in the prompt.

Letting the new developer in

Open the onboarding bullet’s menu and choose Share…. Make an edit link, or under “Or invite somebody by email” type their address and choose Invite them to edit. An edit link opens without an account. An invite becomes theirs the next time they sign in with that address, within thirty days, and Pando does not send the mail, so either way you pass the link on yourself. They need edit to write a question, and edit also lets them change, move and delete anything in the branch; with view they can only read. To take the branch back, open the same sheet, press Revoke on the edit link if you made one, and press Remove on their row under “Who else can see it” if they have one.

Holding Claude Code to that branch

To keep Claude Code out of the rest of your outline, open the onboarding bullet, open Agents and API keys in Settings, tap Claude Code’s row under Your agents, and press Let it reach only “Onboarding: billing-service”, the bullet you are in. It then reads and writes that bullet and everything under it, even if you approved it to read only, and nothing else. The press holds the connector, not the repository: set up with --scope user, Claude Code reaches only this branch in every repository, can no longer reach the memory bullet it had, and remembers here, where the new developer can read and change it. If Claude Code keeps your notes elsewhere in Pando, leave it approved to read and write and skip this.

Two limits

Nothing wakes Claude Code when a question arrives. Pando sends it no notification, so the answers come the next time somebody runs it and asks.
The answers come from your Claude Code, not the new developer’s. Their connector reaches their own outline, and in Pando a grant reaches down a branch, not across into another account, so the branch you lent them is out of its reach.

Where Pando is not the answer

What every session needs. Anthropic’s docs put "build and test commands, coding standards, architectural decisions, naming conventions, and common workflows" in the project CLAUDE.md, shared with "Team members via source control", and /init writes a starting one from the codebase. The branch keeps the new developer’s questions and the answers, and an answer every session needs belongs in CLAUDE.md.
A new developer with Claude Code of their own (a Pro, Max, Team, Enterprise or Console account, or a provider such as Amazon Bedrock) who only wants the overview. Their own Claude Code answers "give me an overview of this codebase", the first prompt in Anthropic’s recipe for someone who has "just joined a new project", in their own terminal, with nothing to connect.

Deeper

Connect an agent, about two minutes · The twelve tools · Which note apps an agent can reach · Who runs this