tibetan-spirit-ops
Six AI agents running a real Shopify business — fulfillment, inventory, customer service, marketing, catalog, finance.
Problem
Running a small e-commerce business is a thousand small operational decisions per week. Founders either drown in the work or hire a team they can’t yet afford. AI tools promise to help but most are point solutions that don’t talk to each other and have no sense of when humans should still be in the loop.
Solution
Tibetan Spirit is a real Shopify D2C business I acquired in late 2025. tibetan-spirit-ops is the AI operations layer behind it. Six specialized agents handle the operational surface. Every customer-facing output goes to a human reviewer first. The system tracks override rates to graduate workflows toward autonomy.
Impact
Processing operational decisions for ~19,400 orders across 559 SKUs. CS drafts cut email response time roughly in half while maintaining brand voice. The graduation framework is the most enterprise-relevant pattern in the portfolio — a working answer to how to move from human-in-the-loop to autonomy.
How it works
Six agents, each with a defined role, model, and per-invocation budget. All run on the same architecture: take input from Shopify or Supabase, draft an output, write it to a task inbox for human review, send a Slack alert. Nothing goes to a customer or modifies a Shopify order without human approval.
| Agent | Budget | Role |
|---|---|---|
| Fulfillment Manager | $2.00 | Order tracking, shipping, supplier coordination |
| Inventory Analyst | $2.00 | Stock monitoring, restock alerts, demand forecasting |
| CS Drafter | $2.00 | Email drafts via triage → enrichment → draft |
| Marketing Strategist | $2.00 | Campaign briefs, content calendar, targeting |
| Catalog Curator | $5.00 | Product descriptions via evaluator-optimizer loop |
| Finance Analyst | $0.50 | Weekly P&L, COGS tracking, margin analysis |
The graduation model — the key idea
Every workflow starts at Tier 3 — Explicit approval. Every agent output goes to a Supabase task inbox. A Slack alert fires. Nothing executes until a human clicks Approve.
A workflow earns promotion to Tier 2 — Draft + 4-hour auto-approve when its override rate (rejected + modified, divided by total reviewed) sits below 25% for 30 consecutive days. Drafts still go to the inbox; if no one rejects within 4 hours, the draft executes.
A workflow earns promotion to Tier 1 — Auto-execute, log only when the override rate approaches zero over an extended period. Execution happens immediately; the inbox entry is informational.
The override rate is the only metric that gates promotion. Not engineer confidence. Not cost savings. Not anecdotal team feedback. Empirical or it doesn’t count.
Hard gates for cultural sensitivity
When AI judgment touches sacred or high-stakes domains, the gate must be deterministic. The cs-triage skill classifies incoming customer emails into seven canonical categories — but spiritual-guidance is checked first, before any other signal. If detected, the email escalates immediately to the spiritual director with no draft attempted.
The same pattern applies to brand voice. Banned terms (exotic, mystical, oriental, ancient secrets) are checked deterministically. Sacred terms that must stay untranslated (mala, thangka, dharma, sangha) are checked deterministically. The model handles the long tail of language; the rules handle the cliff edges.
<HARD-GATE>
Spiritual guidance emails MUST be checked FIRST
before any other classification. If detected,
escalate immediately to spiritual-director —
no draft is created, no automated response
is attempted.
</HARD-GATE>Cost discipline
Every agent has a per-invocation cost budget enforced by a hook that runs before the tool call executes. The hook reads the budget from the agent config, tallies the invocation cost in real time, and blocks the call if it would exceed the cap.
Total monthly Claude API spend across all six agents is tracked in Supabase and visible on the dashboard.
cs-triage-skill.zip
The cs-triage skill + classification matrix. Install into your own Claude Code setup, customize the brand-voice rules for your business.
Includes: cs-triage SKILL.md, classification matrix, brand-voice config
Other projects