Claude partner CI agent: handoff pack
You build, with a teaching Claude Code session, a Claude Managed Agent that sends you a weekly competitive intelligence (CI) debrief on Anthropic's Claude service partners. Goal, in your words: "track where the market is heading and inform our own product positioning." Ray created nothing on Anthropic's side; you approve every resource.
The 5-minute version
- Put this folder in a private GitHub repo, with the seed CSV (comma-separated values file) committed at
data/2026-10-22-index.csv(see "Seed data" below; the cloud session cannot see any Mac's files). Open a Claude Code cloud session on that repo. - Paste
KICKOFF-PROMPT.mdinto the session. - The session builds in 6 phases (
PLAN.md) and stops at each boundary to teach. You approve everyant apply(the command-line interface, or CLI, command that creates resources). - Sequence: one manual debrief, then memory for week-over-week comparison, then delivery, then a Monday schedule.
- First, answer
DECISIONS.md(about 10 minutes).
File map
| File | Use it for |
|---|---|
| [[KICKOFF-PROMPT]] | The text you paste into the cloud session |
| [[PLAN]] | The phased build, with done-when checks |
| [[BRIEF-SPEC]] | What the debrief contains, plus a short CI primer |
| [[RUBRIC]] | Starter grading criteria for the outcome |
| [[DECISIONS]] | Choices only you can make, each with Ray's guess |
Seed data (exists; Playwright scrape of 2026-10-22): 2026-10-22-index.csv, 321 partners (16 Global Premier, 40 Preferred, 257 Select, 8 with no tier badge), with its own index note ([[README]] in claude-partners/, the directory capture and tier-requirements writeup). It lives on Ray's Mac at ~/rdco-vault/01-projects/phdata/competitive-intel/claude-partners/. The cloud session cannot read that Mac, so commit the CSV to the repo at data/2026-10-22-index.csv before pasting the kickoff prompt (ask Ray to push it, or copy it from the vault, then git add, commit and push). Do not rebuild the baseline from the public directory: it is a JavaScript app (the scrape needed about 52 "View more" clicks), and web_fetch will likely return almost nothing.
Order
[[DECISIONS]] -> [[BRIEF-SPEC]] (skim) -> paste [[KICKOFF-PROMPT]] -> follow [[PLAN]] with the session.
Concepts you'll learn
| Phase | Managed Agents concept |
|---|---|
| 0 | Workspace, ant auth login, spend limit |
| 1 | Agent (versioned config file, synced with ant apply), environment (sandbox + network rules), session (one run), outcome + rubric (a separate grader scores, the agent revises), web tool allowlists |
| 2 | Memory store (files that persist across sessions; read-only vs read-write) |
| 3 | Vault credentials (secrets kept outside the sandbox), MCP (Model Context Protocol) servers, permission policies |
| 4 | Scheduled deployment (cron schedule), deployment runs, per-run budget caps |
| 5 | Agent versioning, rubric tuning, cost per run |