Holding ourselves to that promise led us to build something we hadn't planned: a framework for what good human-AI collaboration actually looks like, and a set of principles for keeping our product accountable to it.

What we discovered

We started with a UX audit. Building a real project from scratch, screenshotting every step, annotating every friction point. We compared our results with feedback from our users.

The three main failure modes we identified may be familiar to anyone who has worked closely with an AI-assisted tool:

• Acting before understanding. The Co-designer was treating user input as instructions to execute at maximum scope rather than information to reason from. In a user interview, we heard, “I want to feel more control of the plan.”

• Operating on invisible state. The AI was making consequential decisions without communicating intent and storing chat memory in a hidden location. One user put it plainly, “There were things changing behind the scenes.”

• Broken consent mechanics. The Co-designer had a bad habit of ignoring stop signs. This sounded like, “I didn't tell it to touch the skills!”

What we created

The collaboration matrix

Before designing solutions, we needed to understand the problem structurally. This led us to create the “collaboration matrix,” a 2x2 grid plotting every AI behavior against two dimensions: who initiated it and whether it was reversible.

The four quadrants make the risk distribution visible immediately.

1. Guided is low stakes.

2. Ambient is productive territory including suggestions, recommendations, proactive help.

3. Gated requires good confirmation mechanics.

4. Unilateral is the danger zone, never acceptable. This is where our Co-designer’s most damaging behaviors lived.

Once we could see which behaviors lived in the danger zone, we could make deliberate decisions about each one: redesign the interaction, add a consent gate, or remove the behavior entirely.

The collaborator contract

With the matrix as our guide for the ideal collaboration experience, we wrote what we called the “collaborator contract.” It lays out a set of plain-language principles governing how the Co-designer should behave, written to inform prompting, interaction design, and engineering decisions together.

Our contract covers four areas: goals, obligations, violation signals, and recovery model. The specifics reflect our users and our product. The structure is transferable.

The exercise of writing the contract was as valuable as the output. It forced us to articulate what the user is owed, what the AI is obligated to do, and what a violation looks like on each side. Skipping this step would have led us to fix the surface symptoms without resolving the underlying dynamic.

How we improved our product

But a contract does not change how a tool behaves, design does. For us that meant redesigning every Co-designer UI component from scratch, using the contract as the design brief. Each component is a direct translation of a principle.

Thinking steps exist because no hidden state is a principle. The Co-designer's reasoning is now a transparent, inspectable log written in plain language for a learning designer, not a developer.

Decision points exist because approving one thing is not permission to do the next. Options are now presented as decisions about the work, each with a clear statement of scope, before the AI acts on anything.

Approval flows exist because accept-to-clear is a failure state, not a workflow. Learning goal and course outline proposals now require explicit approval before the AI builds anything.

Why it matters

The quality of AI collaboration is a design problem, not just a model problem. A more capable model does not automatically produce better collaboration. Collaboration requires intentionality: about who holds what role, what each party is owed, and what happens when something goes wrong.

For tools serving non-technical audiences, this is especially consequential. These users bring deep domain expertise and deserve AI that respects it. The Collaborator Contract is how we make sure ours does.

This release expands what custom interactives can do — persistent storage, richer image handling, and a smarter analytics layer, while shipping targeted fixes for the most-reported widget and co-designer bugs.

New Features

• Co-designer model selector - Added a model selector to the co-designer chat panel so you can choose which AI model powers your session. A "New chat" button lets you start fresh without leaving the page.

• Image submissions in Free Response - Students can now attach images in Free Response widgets, and AI grading handles image attachments inline alongside text responses.

• Widget persistent storage - Custom interactive widgets can now store and retrieve user data between sessions using a key-value storage API. Widgets built by the co-designer automatically use these APIs where appropriate.

• Session-level xAPI rollup for custom apps - When a student finishes a custom interactive app, Lazuli now synthesizes the full session into competency-level xAPI statements. Apps that previously emitted only raw events now produce a complete learning story that feeds into analytics dashboards and the inference engine.

• Vimeo URL support - Video panels now accept pasted Vimeo URLs directly, in addition to YouTube and file upload.

• Expanded callout icon picker - The callout icon menu now includes the full Lucide icon library with a search bar, replacing the previous eight-icon palette. The co-designer and the manual picker now draw from the same icon set.

Improvements

• Increased free monthly credits - Doubled free credits from 250 to 500 a month. Based on usage behaviors, we think this will better enable people to evaluate what Lazuli can do and let them make better month over month progress while keeping costs reasonable.

• LTI student identity in analytics - LTI and Lazuli students' xAPI events from custom interactives are now correctly attributed to their real identities; previously they appeared as anonymous actors in the LRS and analytics dashboards.

• AI chat character names - The AI chat widget now correctly reflects the character name and avatar set during design; the sidebar label and widget header stay in sync.

• Student dashboard progress readability - Progress scores now use the standard foreground text color instead of the light green "correct" indicator, making them legible across all themes.

• AI response feedback styling - Polished the thumbs up / thumbs down feedback buttons in the co-designer chat.

• Custom interactive refinements - Fixed styling inconsistencies between controls and image generation widgets. Updated edit-with-prompt to use clear iconography. Improved inline edit prompt dismissal behavior.

• Welcome screen copy - Updated onboarding text on the welcome screen for clarity.

• and multiple Infrastructure and Security improvements including successfully achieving our SOC 2 TI Attestation!
  

  

Bug Fixes

• Document upload failures — Fixed intermittent upload errors that occurred even for files under 1 MB; document uploads now succeed reliably.

• Table rows disappearing after save — Fixed a bug where table rows added with content would vanish after saving a named version and returning to the page.

• AI chat widget page reload — Fixed a bug where the page would unexpectedly reload and delete work in progress while building an AI chat widget.

• Custom interface unintended regeneration — Fixed a bug where updating one element on a page with a custom interactive caused unrelated page content to regenerate.

• Version restore content loss — Fixed page content disappearing when restoring an artifact to a previously saved version.

• Public preview sync issues — Fixed public preview reverting editor changes during collaboration sessions.

• Co-designer AI character description — Fixed the co-designer not properly replacing the AI character's description when building or editing AI chat widgets.

• In-chat slash menu clipped — Fixed the slash command menu in the co-designer chat being cut off by the panel boundary.

• Lightbox image margins — Fixed margin issues when expanding images to full-screen in the lightbox viewer.

• Widget title field corners — Removed the visible border radius that appeared on widget title fields when text was highlighted.

• Component warnings — Surfaced dropped component warnings to agent with syntax hints to prevent silent widget generation failures.

Reach out if you have any questions and much more to come soon!

Released May 14, 2026 · View full changelog

Release date: Wed, Apr 16

This release centered on transforming the Co-designer from an AI assistant into a true collaborator, while simplifying Lazuli's content hierarchy through the Activity-to-Page transition. We completed critical Open edX interoperability work, expanded skills management capabilities, and deepened learning analytics through comprehensive xAPI instrumentation of custom content.

Highlights

Co-designer as a True Collaborator. A comprehensive reimagining of how learning designers interact with Lazuli's AI. The Co-designer now surfaces its thinking process with condensed, non-technical animations that build trust without overwhelming. Added an explicit question-asking tool that prompts for scope and context before taking action—no more guessing what the AI needs. Implemented thumbs up/down feedback on every chat message and aligned all internal tool language with user-facing terminology. Removed the "please continue" suggestion button that had run its course as the AI became more capable.

"Activity" Re-conceptualized—Pages Become the Primary Unit. Moved all activity properties to the page level and eliminated the activity as a distinct structural layer. Skills now assign directly at the page level through the right sidebar's "Page" section. This cleaner hierarchy reduces cognitive load during authoring and aligns the product structure with how learning designers actually conceptualize course organization. No more navigating nested activity/page distinctions—just pages with assignable skills.

Project Memory System. Made project context visible, editable, and user-owned. Project memory now lives in a dedicated location where both users and the AI can update it dynamically as projects evolve. Context is no longer hidden in chat history—it's a first-class artifact that learning designers control directly while benefiting from AI assistance in keeping it current.

Skills Management Suite. Complementary capabilities that open skills to customization and portability: upload custom skills as .zip files through a lightweight interface (no in-platform authoring required), convert existing agent instructions into the unified skills framework, and download skills for backup or cross-environment sharing. Together these changes position skills as portable, user-controlled building blocks for specialized authoring behaviors.

OLX Export Finalization. Completed the standalone HTML generation and API endpoint work for OLX exports. Style handling for assessment problems now works seamlessly in exported archives. Experimented with an OLX Export MCP tool for programmatic access to the export pipeline, which can enable integration with external workflows and CI/CD systems.

xAPI from Custom Interactives. Instrumented custom interactive widgets to emit xAPI learning statements, extending analytics coverage beyond built-in assessments to any AI-generated learning activity. Whether created in native Lazuli or viewed through an MCP app such as in Claude Code, custom interactives now generate structured learning data that feeds into dashboards and inference engines.

HTML Native Artifacts. Introduced support for HTML-based reports and artifacts, expanding beyond the widget ecosystem for custom dashboards and data visualizations. Learning designers can now create rich, interactive documents that live alongside course content.

Math Formula Editor Brand Alignment. Replaced the legacy alert-box math editor with an inline interface that matches Lazuli's design system. Supports visual formatting (font, color) for both inline and block-level expressions, making mathematical content feel native rather than tacked on.

Improvements

Page Headers Deprecated. Removed the old page header system in favor of simpler title options and full-width image support. Content gets more breathing room, visual hierarchy feels cleaner.

LTI Deep Link Streamlining. Removed the review screen from the deep link picker and simplified course player link selection for LMS integrations.

Widget Testing with Synthetic Students. Added the ability to test open response and conversation widgets with "good" and "bad" synthetic student personas, enabling validation of assessment design before publishing.

Widget & Editor Polish. Fixed sticky resize handles on custom interactives and iframes. Added image linking functionality for pseudo-button behavior. Improved media widget icon differentiation so images/videos/audio don't all look identical in the sidebar. Fixed AI chat character images not displaying. Resolved horizontal line drag behavior in callout boxes, dark mode display inconsistencies, and footnote reference scrolling that went behind the sidebar.

Sidebar & Publishing. Long filenames now truncate properly in upload modals instead of breaking layout. Only content in the "Course" folder appears in publish dialogs—Plan, Library, and Dashboard folders correctly excluded.

UI Cleanup. Removed the deprecated Evaluation tab from the right sidebar and broken custom properties field from the skills graph. Fixed invitation email branding inconsistencies.

Security Hardening. Completed penetration test remediations: session hijacking mitigations, security headers (CSP, HSTS, X-Frame-Options), and EXIF metadata sanitization on uploaded images. Post-CSP security fixes deployed.

What's Next?

Over the next month, we'll be:

• Increasing reliability of the Co-Designer

• Making widget generation more consistent and stable

• Improving the measurability of interactions which is the foundation of enabling valuable course insights

• Strengthening embeddability and interoperability with other platforms

Dr Porcaro was invited to speak at the US Department of State sponsored workshop Corporate Governance for Resilient Investment in the Era of Digital Transformation, hosted by World Learning as part of the Leaders Lead On-Demand project. His presentation "The Promise and Peril of Agent-Based Innovation in Regulated Sectors" discussed the advances of AI coding agents and especially the possibilities made possible for small and medium-sized enterprises across Central Asia with Agent Skills, as well as some of the areas of continued need for governance and policy. He shared how tools like agent skills allow small organizations like Lazuli with domain experts pared with focused teams of designers and developers can have massive impact on their fields.

It was a great opportunity to meet with government, business, education and non-profit leaders from across Kazakhstan and Kyrgyzstan, and to spend some time experiencing the digtal leadership of the beautiful city of Almaty. We can't wait to continue conversations started and to some day return to this region

Release date: Wed, Mar 19

This release was focused on deepening Open edX interoperability, strengthening organizational customization, and improving the authoring experience with refined widget controls and refined assessment interactions.

Highlights

OLX export & Open edX integration. A landmark feature for our Open edX interoperability roadmap: courses now export as OLX .tar.gz archives that deploy directly to any Open edX platform, preserving Lazuli functionality including competency tracking, xAPI capture, and embedded interactive content. This enables Open edX institutions to host Lazuli content natively within their infrastructure.

Organizational branding customization. Expanded the brand customization system with organization-level branding controls, allowing administrators to set unified design guidelines across all projects and gives organizations control over their learning environment's visual identity.

Custom interactive styling control. Added HTML/CSS editing controls to custom interactive widgets, enabling fine-grained style adjustments, including color choices for prompts and direct HTML editing. Learning designers can now brand interactive content to match organizational standards without leaving Lazuli.

Activity to Page Hierarchy Simplification. Deprecated the activity as a unit of design and moved all activity-level properties to the page level. Removed activity sections and now manage skills and properties through the page-level "Page" section in the right sidebar. This cleaner hierarchy makes skill assignment and content management more intuitive and reduces unnecessary structural complexity.

Skills map & course structure decoupling. Separated skills map generation from course structure building, enabling independent updates to both systems. Teams can now collaborate with the Co-designer to create course structures in which pages have a one to many relationship with skills.

Improvements

Assessment & widget refinements. Multiple UX improvements to assessment components including MCQ radio button alignment fixes, consistent styling for free response text boxes, media widget default caption behavior, and improved carousel slide height consistency. Custom interactive replacement now includes a back button for cancellation.

Editor enhancements. Resolved sidebar behavior for better document organization. Delete key functionality improved to properly remove selected widgets without character-level deletion.

Grading service expansion. Extended grading service as an external API, supporting large video file submissions (up to 2GB) and reducing infrastructure constraints.

MCP tools & integration foundation. Established groundwork for Lazuli as an MCP tool, expanding integration capabilities with Claude and other AI platforms.

Image Captioning. Moved image caption controls to the right sidebar for consistency with other widget editing patterns, reducing cognitive load during content authoring.

What's Next?

Next release we're diving deeper into evaluation and analytics, improving how our Co-designer collaborates with learning designers, and giving the organization homepage a refresh.

We hope you enjoy the updates and, as always, don’t hesitate to send along any feedback!

From a learning standpoint, it improved several things. Instead of passively scanning questions with answers right beside them, she actively engaged with each one. Questions were delivered in ten-question chunks, keeping sessions focused and manageable. And because the format matched the actual competition—identify the book from a question and choose from six titles—the practice felt authentic.

Most importantly, she had real agency. She could see her results instantly, retry missed questions, and track her progress session by session.

Ten minutes of setup. A study tool she actually used. That's the Lazuli sweet spot.

Release date: Thu, Feb 19

This release was all about making the document editor more intuitive and powerful, with the new widget editing approach being the standout feature that fundamentally changes how you work with interactive content in Lazuli.

Highlights

Streamlined widget editing. The biggest change this release is a complete overhaul of how you edit widgets in the document editor. We've removed the floating edit bar and moved all widget configuration to the right sidebar. You can now make changes directly to content within widgets without entering a separate edit mode, making the editing experience faster and more intuitive.

Deep linking improvements. You can now deep link the Lazuli course player in your LMS, making it easier to integrate Lazuli content directly into your learning management system workflows.

Document editor refinements. Added undo and redo buttons to the document editor, giving you more control over your editing workflow. These buttons make it easy to quickly step backward or forward through your recent changes.

Improved custom interactive generation. Significantly enhanced the Custom Interactive widget (previously called "Generative UI") with a more intuitive generation experience and better overall usability.

Export flexibility. You can now export individual widgets, activities, or full pages, giving you more granular control over what content you share or export from Lazuli.

Improvements

Security enhancements. Completed a comprehensive API security refactor to standardize access control across the platform.

Bug fixes. Resolved various issues including fixes for AI chat editing, sidebar file deletion, and student experience improvements.

Engineering improvements. Enhanced testing infrastructure and improved code quality across the codebase.

What's Next?

Next release we're focusing on organizational branding customization and course evaluation tools. We're also continuing to improve Open edX interoperability and refining how skills maps work within course planning.