AI Design and Migration Assistant
A multi-mode assistant that helps learning and development teams migrate and design courses on a new AI-first platform, with governance built into every step.
A migration with no room for error
Multiple teams in a learning and development organization had under a year to move a large content library onto an AI-first LMS that did not ingest existing formats cleanly and lacked clear documentation on how it functioned. The work forked immediately. Compliance-critical content had to transfer with exact fidelity, since any change would trigger a fresh legal review the teams had no capacity to absorb. Older content needed substantive improvement or updating before it moved. Because the platform was lightly documented, even routine builds took trial and error to get right. No one had been assigned to solve the underlying problem, so I took it on.
Prioritize by risk, then build for reuse
I sequenced the work by business risk and built the compliance-critical path first, where the risk was highest. Rather than solve the problem only for my own work, I designed the assistant for every team touching the platform and packaged the hardest-won knowledge into resources others could run on their own. That turned a personal workaround into shared capability, and it meant the governance had to hold up across many hands, not just mine. I used the ISO/IEC 42001 AI governance standard as a framework to ensure the tool held up for appropriate and ethical enterprise use.
A multi-file system with three modes
At the time of creation, the assistant was built as a multi-file AI project rather than a single prompt due to complexity, separating concerns and keeping each part inspectable. A dedicated project file governs the system’s gates and guardrails to prevent data drift. It runs in three modes and produces a complete, execution-ready build package as its output.
Designs courses from a brief or outline, applying instructional standards and the platform’s native components from the start.
Moves existing content across exactly, changing structure only where the platform technically requires it. Built for content that must stay verbatim for compliance.
Carries existing content across while upgrading it, adding the structure, assessments, and standards the original lacked.
Captures every source unit as an immutable baseline before conversion begins, so nothing can be lost or reordered without detection.
Separates what the platform ingests from what people read, so build instructions never leak into learner-facing content.
Checks all content against original source files to ensure no substantive changes in conversion modes. A second gate for high-fidelity mode guarantees 100% content accuracy. Conversion Plus mode incorporates designer input on changes and produces a changelog for versioning.
Two resources the whole org can use
Two parts of the system were built to stand on their own, so colleagues could benefit without adopting the full assistant. The first is a platform-findings reference: a living knowledge base of how the platform actually behaves, with every finding marked as tested or assumed and dated, written to hand to anyone working with the same platform. The second is a drop-in capability file, a single markdown file a colleague adds to their own AI project to instantly inherit fluency with the platform, including how to prompt each feature and learning and design considerations. The first lowers risk across the org. The second shortens the learning curve for every team without requiring my system.
Accountability built in from the start
Governance was not bolted on at the end. It shaped the architecture from the first version, and four mechanisms carry most of the weight.
Decisions on high-stakes content (compliance, legal, financial) route to a human reviewer by rule, and the assistant is barred from inventing or correcting those facts on its own.
A build-blocking reconciliation step measures the converted course against the frozen baseline and confirms nothing was lost or silently reordered before anything ships.
Platform behavior is documented by confidence level and dated, so no one acts on an unverified assumption, and untested behavior is probed rather than trusted.
The architecture separates the working model from platform input from human-readable output, which keeps build instructions from ever surfacing as learner content.
From discovery to a reusable system
Reverse-engineered the undocumented platform through structured A/B testing, isolating not just single settings but the interactions between them.
Designed the three modes, the source-inventory baseline, and the two-artifact output model, keeping each part inspectable.
Built the compliance-critical path first, where the cost of error was highest, then layered the improvement and design modes on top.
Wired in the guardrails, the reconciliation gates, and the human-in-the-loop rules as the system grew, rather than retrofitting them.
Extracted the hardest-won platform knowledge into standalone resources any team could run on its own.
AI Governance · Operationalizing AI · Prompt Engineering · Meta-Prompting · Model-Aware Design · Risk Management · Business Process Management · Workflow Automation · Human-in-the-Loop Design · ISO/IEC 42001 Alignment · Instructional Design · Capability Building
