Back to library
AutomationOperator Vault

AI Agent SOP Builder Pack

Turn a messy repeatable process into a safe, testable AI-agent operating procedure.

Runtime
60-90 minutes
Difficulty
Advanced
Format
Automation design pack

Example run

The artifact is the proof.

Tested 2026-07-12 · Claude (Fable 5)

Scenario: Refund-desk SOP for a 3-person Shopify candle store (~900 orders/month) whose owner wants inbound refund emails triaged and settled without double refunds or policy leaks.

Sample artifact

From the example run
DECISION TREE (first match wins) — 1: ticket contains chargeback, dispute, attorney, fraud → hard stop, tag dispute, no reply. 2: no order number found → send order-number template, tag pending-info. 3: order older than 30 days or final-sale → draft decline reply dry-run, ask and wait. 4: amount ≤ $100, reason in damaged / wrong item / tracking-shows-lost → queue refund for one-click approval, confirm only after refund verified. 5: amount > $100 → ask and wait. 6: otherwise → model classifies and drafts dry-run, notify async below 0.7 confidence.

Before

A Gorgias inbox, a 30-day policy doc, and a plan to 'let the agent handle refunds' — no thresholds, no gates, and no answer for what happens when the refund API hangs mid-payout.

After

A nine-section SOP: a 6-rule decision tree with model judgment only as the fall-through, all six side-effect classes gated (deletes and off-platform payouts forbidden, refunds approval-queued because money is irreversible), a filled escalation matrix, a refund_action JSON contract, and five acceptance tests — the duplicate test reproduces the store's actual past incident, a double refund on a re-opened ticket.

Run it like this

  1. 01Fill the six-line inputs block — be literal in the two action lists
  2. 02Paste the master prompt section plus your inputs into a fresh session
  3. 03Fill every escalation cell with a number or keyword list — the prompt rejects 'when appropriate'
  4. 04Classify each allowed action reversible, compensable, or irreversible; approval-gate the irreversible ones
  5. 05Run the five acceptance tests before real data; on any fail, fix the SOP, not the test

What you leave with

Useful output, not another chat.

01

Nine-section agent SOP with a decision tree

02

Gated tool policy and escalation matrix with concrete thresholds

03

Five acceptance tests you run before launch

Contents

The working parts.

Best for Operators building internal assistants, lead routers, support agents, or JARVIS-style workflows.

Master prompt
Side-effect gate checklist
Escalation matrix template
Output contract template
Acceptance test suite
Worked example: lead-triage SOP

Useful for

Internal agentsLead routingOps automation