Prompt-centric framing
- Every ad starts from scratch
- Claims live inside prose
- Review arrives at the end
- Output count looks like progress
01 · Positioning
I design the operating systems around AI.
The workflows, evidence, controls and human decisions that turn model capability into a service a team can actually use.
A governed activation workflow for a regional insurer.
Product strategyService designGovernanceInformation architectureAI prototyping
02 · The reframe
The request was “more variants.” The real problem was controlled decisions.
→
Operating-model framing
- Reusable brief and claim context
- Locked facts; flexible expression
- Human gates at decision points
- Accepted packages define value
Generation speed matters only if strategy, evidence and approval can absorb the output.
03 · The system
A repeatable path from brief to reviewable activation.
Select a stage to see what the workflow carries forward.
Structured brief
Settles the audience, funnel stage, product/state context, placement and messaging territory before generation expands.
04 · Information architecture
Four kinds of authority. None can impersonate another.
Is it appropriate?
The brief and messaging territory establish where the work should go.
Is it supportable?
Claims, citations, restrictions and applicability establish what may be said.
Does it work?
Hooks, transitions and executions require human creative judgment.
Is this accepted?
An authorized person accepts the exact version for a defined use.
The model can propose. The system constrains. Authorized humans decide.
05 · What I built
Not a prompt library. An executable service foundation.
Strategy layer
Messaging matrix, literal catalog and analysis kept direction separate from authority.
Claims layer
A seeded registry and draft bundles modeled wording, evidence, use and restrictions.
Assembly layer
Packet retrieval and structured copy scaffolds connected source fields to drafts.
Validation layer
Python checks, schemas and stable failure codes made narrow rules testable.
Human review layer
Four gates covered brief readiness, messaging, line-level copy and final assets.
Operating layer
Sprint templates, governance, reusable skills and evaluation records defined the handoffs.
06 · Evidence-led validation
I turn uncertainty into engineering decisions.
4 / 4Supplied dry-run fixtures reproduced their expected outcomes.
10 / 10Starter golden cases reproduced five intended passes and five intended failures.
19Targeted probes mapped control boundaries and exposed edge conditions early.
1Repeatability risk surfaced early and became an explicit hardening requirement.
The result: a working foundation, a defensible evidence trail and a concrete path from prototype to dependable operation.
07 · Bounded-agent roadmap
Make the workflow authoritative. Keep the model bounded.
Model may
- Propose angles
- Draft flexible blocks
- Explain failures
- Prepare handoffs
Deterministic control kernel
- Evaluate claim eligibility
- Insert locked content
- Enforce schemas and dependencies
- Own states, versions and retry rules
- Bind decisions to exact artifacts
Model may not
- Edit claim authority
- Approve its own work
- Override escalation
- Publish or traffic ads
Humans retain strategy, legal, creative and final-asset decisions.
08 · Commercial thinking
Optimize for accepted throughput, not generated volume.
PRIMARY UNIT
Cost per accepted asset = attributable labor + tools + production + QA + rework + setup allocation ÷ accepted assets
First-pass technical acceptance
First-pass human acceptance
Review minutes per package
Rework by failure class
Evidence completeness
Changes after approval
The commercial offer is a controlled production service with explicit scope, traceable decisions and accepted outputs—not “unlimited AI ads.”
09 · Why hire me
I connect strategy, systems and proof.
For product teams
I turn ambiguous AI opportunities into explicit workflows, contracts and evidence gates.
For agencies
I design repeatable services that protect judgment while reducing coordination overhead.
For AI builders
I bridge prototypes and dependable operations through schemas, states, controls and evaluation.
My pattern: identify the real constraint, make the information model explicit, build the smallest testable system, then define what evidence earns the right to scale.