Copy this into Claude Code
Prompt
Or onboard yourself
Agent playbook
This section is the contract a coding agent follows. People can skip to Troubleshooting. Two steps need the person: approving the browser sign-in, and restarting Claude Code so the tools load. Don’t try to work around either. Do everything else yourself.1
Preflight
2
Check the setup
ready: truemeans signed in with Claude Code connected: go to step 5.workspace: null,server: "revoked"or anything missing: go to step 3.server: "down": tell the user their server isn’t answering, show them the report’snextadvice, and stop.
3
Sign in
/os-* commands. Nothing is written into any repo. Login ends by printing the same report as step 2, so you don’t need to run status again.4
Restart Claude Code
Claude Code loads OpenScaffold’s tools when a session starts. Tell the user, verbatim: “Restart Claude Code (exit, then run
claude again) and say ‘continue onboarding’.” Then stop.5
Confirm the tools
Call the
list_deployments tool. If it isn’t available, the session wasn’t restarted: repeat step 4.If the customer already has a deployment, call use_deployment and don’t create another. If its harness_version is set, it already has a plan: call get_briefing, show it, and you’re done. If it’s null, go to step 6 but skip create_deployment: just add the customer’s material.6
Create the first deployment
Ask the user for the customer’s name, and which files or folder hold their material: call transcripts, ticket exports, email threads, docs.Call
create_deployment with the name (unless step 5 found one), then add_document once per file, with its type, title and text. Only add files the user named. If their team has connected Gmail, Slack or other tools, tell them they can feed this deployment from Sources at app.openscaffold.ai instead.7
Build the plan and show the briefing
Call
generate_harness. If it says documents are still being read, wait a minute and call it again. It runs for a minute or two. Call get_briefing once a minute, at most five times, until it shows a plan.If it still says “No plan yet”, generation stopped: call generate_harness once more and wait the same way. If that fails too, tell the user it didn’t finish and to check the deployment’s Activity at app.openscaffold.ai; don’t keep retrying.Once the plan exists, show the user the briefing: the goal, the must-haves, the rules and the open questions, each with where it came from.Rules
- Never invent the customer’s name, requirements or files. Ask.
- One deployment per customer engagement. If it already exists, use it; don’t create a second.
- Only add files the user named. Never upload anything else from their machine.
- Every command here is safe to re-run. When unsure where you are, run
npx -y @openscaffold/cli status --jsonagain. - Don’t edit
~/.claudeor credentials by hand, and don’t call internal APIs.
Troubleshooting
Status says the server isn't answering
Status says the server isn't answering
The login points at a server that’s down, often a local dev server. Start it, or sign in to the hosted app instead (a plain
login reuses the saved server):The login link expired
The login link expired
Links last 10 minutes and work once. Run
npx @openscaffold/cli login again.The tools aren't there after login
The tools aren't there after login
Claude Code loads them at session start. Restart it and continue.
Claude Code isn't on PATH
Claude Code isn't on PATH
Login prints the exact
claude mcp add command to run yourself; status shows the MCP server as unknown until then."No findings yet" from generate_harness
"No findings yet" from generate_harness
The documents had nothing the plan could use, or were empty. Add the customer’s calls, tickets or emails and try again.