ID Assistant

Challenge

An instructional design team spent review time on work a model could draft: QA-lead rubric checks and subject matter expert reviews of training material, video request scoping, first-pass scripts, and caption files. The tool had to live in Microsoft Teams, where the team works, with no admin rights and managed hardware. Output feeds real training, so an invented finding or message was worse than none.

Process

ID Assistant is a two-model system: Claude Code, connected to Microsoft 365 through MCP connectors, authors and maintains it; Copilot Studio runs it live in Teams and Microsoft 365 Copilot. It grew out of a v1 file-drop job system on headless Claude, with connected tools now hiding the underlying folders.

Six specialist tools are in production. QA Review and SME Review apply the team’s rubrics, quoting the passage behind each finding so fabrications are detectable. Plan a Video turns requests into scoped briefs, flagging weak answers instead of filling them in. Write a Script produces a script and formatted Word document from a locked template. Caption Workbench pulls transcripts from SharePoint and Stream and writes caption files back, with no paid transcription service. Submit Feedback routes bug reports to a SharePoint list and Teams channel.

Measurement drove hallucination control: an evaluation run caught the brief tool inventing tone, messaging, and a date on one-line requests. Thin requests now trigger questions, a source-material check blocks invented pain points, and an automated evaluation runner with a routing test set measures tool selection. The planner ignores what it can, so hard rules live in flow logic and prompt text, not tool descriptions.

The agent is low-code, run as code: the Power Platform CLI exports it as YAML into Git, and Python generators produce flows, prompts, and Word templates. Deployment safeguards prevent mishaps: a fresh-clone diff before every push, generated artifacts as source of truth, and a post-push landing check. A tool returning HTTP 200 on failure prompted a project-wide rule that every flow ends in explicit success or failure. When Microsoft’s API rejected hand-built Word content controls with a misleading error, I diffed against one Word-authored control, found the missing parts plus three undocumented rules, and built a template generator with Word as the final verification gate.

Result

Shipped through Teams and Microsoft 365 Copilot in September 2026, now in daily use. It has produced real training and product-launch video scripts from meeting transcripts, decks, and messaging documents, every claim cited to its source. Built solo in a locked-down Fortune 500 environment. More on the resume.