marketing-agent-team
A one-person marketing team of AI employees, modeled on solomi's "Series AI Agent" (TikTok @_solouniverse), running hosted on Anthropic Managed Agents. Nothing runs on this Mac: this folder is the source code and config that gets synced to the platform and, from phase 2, deployed as a web app.
Spec (final v1): https://claude.ai/artifact/QPnbQhz1UXPs6nY5vqMHYG
Phase 1 (this): the team as files
| Path | What |
|---|
agents/lead.md | The coordinator (điều phối). Plans each week (goal, post mix, one test, plans/-plan.md), picks research's topics to fit it, runs the weekly cycle, delegates, reports. Roster: the four employees below. |
agents/research.md | Research employee. The only one with web search and fetch. Writes research/. |
agents/writer.md | Writer. Drafts into content/drafts/. No web, no bash. |
agents/qa.md | QA-QC. Sets status: ready or needs-fix plus a checklist. No web, no bash. |
agents/metrics.md | Metrics. Turns a CSV in reports/ into reports/<week>.md. |
agents/scout.md | Skill scout, outside the team (not in the lead's roster). Searches the web and proposes one skill per run for the boss to approve; the web app starts it and delivers approved skills into memory. |
environments/cloud.yaml | The sandbox container (Anthropic cloud, unrestricted egress for now). |
HOUSE_RULES.md | The eight nội quy. Master copy; each agent prompt repeats the ones that bind it. |
brand/BRAND.md | The brand card template. You do not have to fill it: scripts/new-brand.sh "<one sentence>" makes the team write it. A filled copy here is mounted on start instead. |
prompts/ | The first-run and weekly kickoff messages. |
scripts/ | start-office.sh, new-brand.sh, update-brand.sh, run-week.sh, say.sh, watch.sh (see below). |
claude-lock.json | Written by ant apply; maps each file to its platform ID. Commit it. |
Agent frontmatter = the Managed Agents agent config; the Markdown body = its system prompt. Edit a file, run ant apply again, and the agent gets a new version. A running office keeps the version it opened with; the web app sends the lead's new text along with that office's next order (workers are not updated this way).
Run it
brew install anthropics/tap/ant # once
ant auth login # once; opens a browser, no API key to store
ant apply --dry-run -v agents environments # read the plan
ant apply agents environments # create the 5 agents and the environment
scripts/start-office.sh # one session = the office; sends first-run
scripts/new-brand.sh "bán matcha cho dân văn phòng, đăng Threads" # the team writes brand/BRAND.md itself
scripts/watch.sh # stream what the team is doing (Ctrl-C to stop)
scripts/run-week.sh "tuần này tập trung vào matcha" # Monday
scripts/say.sh "approve content/drafts/2026-11-02-matcha-sang.md"
Outputs land in /mnt/session/outputs/ inside the session and show up on the office page (web/). scripts/show.sh content/drafts prints them in the terminal instead.
Decisions
- One long-lived session is the office. The sandbox keeps
research/, content/, reports/ between weeks. BUDGET_CENTS (default 1000 = 10 USD) is a hard list-price cap on the session; raise it with ant beta:sessions update when it pauses with budget_reached.
- Models: lead on Claude Opus 5.5, employees on Claude Sonnet 5.5. Change in the frontmatter.
- Permissions: file tools under the platform's
auto policy; web only on research; bash off for writer and QA. There are no sending tools at all in phase 1, so rules 1 and 3 cannot be broken yet. Phase 3 adds Gmail/Notion/Drive via MCP with always_ask on anything that sends.
- Networking:
unrestricted because research needs the open web. Tighten to limited + allowed_hosts once the competitor list is stable.
- Language: prompts in English with the Vietnamese terms the creator uses; the team writes content in Vietnamese per BRAND.md.
Phase 2: the office page
web/ is a Next.js app: dashboard, drafts with Approve / Reject, research, reports, brand, chat. See web/README.md to run it locally (cd web && pnpm dev) and to deploy it on Vercel with Supabase.
The lead copies every file it creates or changes to /mnt/session/outputs/; the page pulls those through the Files API, so nothing in the sandbox is ever reached directly.
Next phases
- Gmail, Notion and Drive through MCP, credentials in a vault, approval gate on every sending tool; weekly schedule as a deployment.
- Design, Publish and Reply employees.
