Business ownership, approval authority, source accounts, data rights, and final operating decisions.
Know what you own before any tool acts.
You keep the accounts, data, approval rights, and operating history. Vendors provide replaceable infrastructure, and the workflow makes every boundary visible.
Source records, Atlas bindings, actions, approvals, runs, receipts, and operator views stay in the owned system.
Hosted services, APIs, uptime, product limits, and platform-specific capabilities.
You should be able to leave with everything that matters.
The stack becomes useful when you can trace the service path and see where work stops. It shows what you keep and what each external platform provides.
The stack should read like a handoff, not a vendor diagram.
A non-technical team can follow one service path: map the boundary, pilot one safe workflow, then control risky actions only when live work needs it.
Name the workflow, source accounts, decision owner, first action boundary, and evidence needed before delegation.
Turn one repeated handoff into scoped actions, durable state, receipts, and a runbook only after the safe path is clear.
Classify live actions as auto-allowed, approval-needed, or blocked with a reason before the workflow touches risk.
The workflow boundary decides what tools are allowed to do.
Source data, assistive work, human judgment, stop conditions, and the audit trail belong on one operating map.
The workflow boundary decides what tools are allowed to do.
This read-only map shows the stack promise in workflow terms. Tools can route and prepare, but anything consequential stops for named authority.
1. Map Claim intake packet before execution. Claims operations owner owns the operating path. The canvas makes the workflow, handoffs, and next decision legible before an agent or system acts. 2. Claim system triage can run when the rule is clear. Claim system triage coordinates with Coverage context summary; the map keeps AI assistance bounded to the work it can safely support. 3. Adjuster review stays with a person. Adjuster handles coverage uncertainty, fraud flags, payout, denial, and sensitive messaging. 4. No payout or denial authority is the stop condition. Stop before coverage decision, payout, denial, or fraud escalation without authority. 5. Claims queue receipt shows the receipt. Shows triage class, missing documents, adjuster state, and customer-contact trail. 6. Use the map as booking context for a workflow pilot. The map has enough owner, assistive work, system behavior, and decision context for a first run.
- allowed routing
- named authority
- stop conditions
- audit trail
You keep the receipts, not a mystery stack.
The technical stack can change. The durable asset is the workflow boundary: source accounts, scoped access, allowed actions, stop states, approval owners, runbooks, revocation paths, and evidence.
One workflow, source systems, owners, handoffs, and failure points.
What your team owns, what CREATE SOMETHING owns, and what vendors provide.
Tools, resources, auth scopes, allowed actions, and transport limits.
Auto-allow, approval-needed, and blocked states with reasons.
Recovery, release evidence, rollback notes, and operator handoff.
The visible state in Substrate or a client-owned application.
Model-opinionated in practice. Model-portable by design.
CREATE SOMETHING builds primarily with OpenAI Codex. The durable client asset is data, MCP contracts, harnesses, skills, prompts, policy, evals, receipts, routing, fallback, and recovery—not access to one model. CREATE SOMETHING owns the system. Cloudflare provides infrastructure. OpenAI provides intelligence. Substrate is the owned database and operator layer; it keeps workflow state, policy, and receipts under CREATE SOMETHING control.
Current instrument and portable exit
The primary environment for setup, demonstration, repository work, and agent-operable delivery.
The workflow boundary, organizational context, approval rules, evals, receipts, and recovery path stay inspectable and portable.
The same contracts and golden tasks can evaluate Claude, compatible harnesses, open-weight executors, and custom models.
Owned system and platform roles
The CREATE SOMETHING database and operator layer owns source records, Atlas bindings, workflow actions, approvals, runs, receipts, and API/MCP access.
CloudflareWorkers, D1, Durable Objects, queues, and edge routes keep the workflow deployable while account ownership, billing, and rollback evidence stay explicit.
The primary reasoning and agent environment stays surrounded by scoped tools, approval behavior, evals, and traceable context instead of hidden authority.
Bring the workflow, the accounts, and the decision owner.
CREATE SOMETHING maps the stack boundary and defines the first controlled path. Before implementation starts, you can see what tools may do and what stays visible.
- Owner
- Workflow owner
- Authority
- Owned stack boundary
- Proof
- Controlled path + decision states
- State
- review
- 01 / Workflow First workflow map
Objects, source systems, owners, handoffs, and failure points.
- 02 / Boundary Vendor and ownership boundary
What your team owns, what I deliver, and what vendors provide.
- 03 / Control Decision states
Auto-allow, approval-needed, and blocked states with reasons.