AI Studio | Andressa Horta

RoleProduct owner and designer

StagePrototype and technical foundation

Core loopWatch, Try, Share

The short version

AI Studio is a peer learning platform where colleagues share short videos of how they actually do their work. Other people watch, try the skill that same week, and share what happened. AI handles the tedious packaging (transcripts, summaries, skill tags, a practice activity), and a human reviews everything before it goes live.

Before and after: this started as a peer-learning community concept and is now being developed as AI Studio, a product.

The problem

Most of what makes a team good at its job never gets written down. It lives in how a senior seller handles an objection, how an architect turns vague requirements into a design, how one person uses AI to shrink an afternoon of research into twenty minutes.

Formal courses cannot keep pace with that. By the time a course is built, the practice has moved on, and a lot of it never becomes a course at all. AI widens the gap: people are experimenting on their own and finding things that work, but nobody else sees it. The best prompt on your team is probably sitting in one person's chat history right now.

How it works: Watch, Try, Share

Step 1WatchA colleague shows a real task in 3 to 7 minutes and explains each decision, including the mistakes. Viewers get an AI-generated summary, key takeaways, chapters, and a searchable transcript.

Step 2TryEvery video comes with one practical activity and a reflection question. The learner does it on real work and records what happened.

Step 3ShareLearners react, ask the contributor questions, post application stories, and can answer with their own version of the demonstration.

Around the loop

CommunitiesContent is organized by topic, such as Customer Discovery, Leadership, AI at Work, Solution Design, Sales Excellence, and Product Knowledge. Each has a lead and a weekly challenge.

Demonstration requestsAnyone can ask for a video they need. Others vote, follow, or volunteer to make it.

Prompt LabPeople post a prompt and the result it produced, remix each other's prompts, and vote on what impressed them.

AI Learning MapA few concrete questions about someone's role, goals, and real AI experience show what they have mastered, what to learn next, and why.

Review and governanceAI flags risks with evidence and timestamps, and an administrator makes the final call.

Privacy by defaultMembers appear as initials only, so nicknames and privacy are protected.

The AI skills map

The Studio organizes examples against a map of AI skills and subskills, so every submission can be tagged, searched, and practiced at a level. This is a draft map of 33 subskills across 8 skill areas. Drag to rotate it, or use the list below.

Skill areas in color, subskills around them. Drag to rotate.

View the skills and subskills as a list

AI foundations (Engage)
- How AI works
- Capabilities and limits
- Spotting AI in everyday tools
- Shared vocabulary
Prompting (Description)
- Clear instructions
- Context and role
- Examples and formats
- Iterating and refining
- Reusable prompt templates
Delegation (Delegation)
- Choosing the right task
- Human and AI split
- Automate or augment
- Mapping a workflow
Evaluation (Discernment)
- Checking facts and sources
- Spotting bias
- Setting quality criteria
- Testing outputs
Responsible use (Diligence)
- Data privacy
- Ethics and IP
- Transparency
- Policy awareness
Tools and agents (Manage)
- Choosing a tool
- Connecting data and apps
- Workflow automation
- Building and supervising agents
Creating with AI (Create)
- Ideation
- Drafting and editing
- Analysis and synthesis
- Multimodal work
Shaping AI (Shape)
- Testing and improving systems
- Data quality
- Feedback loops
- Contributing to governance

Levels, shown for Prompting

Each subskill can be described at levels. This is an illustrative example of the idea; the level wording is a draft to validate with subject experts.

Level 1AwareWrites simple requests and reads the answer.

Level 2PractitionerAdds context, role, and the format wanted; revises when the answer misses.

Level 3AdvancedUses examples, breaks big tasks into steps, and keeps reusable templates.

Level 4LeaderBuilds shared prompt libraries, tests them against real cases, and coaches others.

The map is shaped by public frameworks, not copied from any of them: the European Commission and OECD AI Literacy Framework (Engage, Create, Manage, Shape) and the AI Fluency framework by Rick Dakan, Joseph Feller and Anthropic (Delegation, Description, Discernment, Diligence). The skill areas, subskills, and levels are a working draft by Andressa Horta.

Where it stands today

What existsA clickable prototype covering every flow above, including contributor upload, AI enrichment, review, and moderation. Also a deployable web app foundation with a database schema, email sign-in, a video publishing adapter, and an AI enrichment adapter.

What does not exist yetReal users, real content, and measured results. That is what a pilot is for, and it is better to say so plainly than to invent numbers.

Why it matters to the business

It captures expertise before it walks out the doorAttrition, reorgs, and retirements take working knowledge with them. A three-minute demonstration from the person who does the job is a far better record than a slide deck about it.

It teaches AI on real workAdoption stalls when training is generic. People believe a colleague who says "here is how I used it on my last project" far more than a vendor demo. The weekly Try step turns watching into doing.

It is cheap to produceA contributor records a screen and talks. AI does the transcript, summary, tags, and practice activity, which beats the cost and lead time of building a course.

It shows what people need to learnEvery request and vote is a demand signal, so learning teams see which skills people ask for and which nobody has answered yet.

It is governed from the startNothing publishes without human review, and AI only recommends. Contributors confirm consent and confidentiality at upload, and repeated violations follow a clear path from warning to suspension to ban.

It builds on what you already ownVideo hosting runs through an existing video platform, so there is no new video system to buy or secure.

How it is different

	Traditional LMS	Video library	Team chat	AI Studio
Who creates content	L&D	L&D or vendors	Anyone, but it disappears	Colleagues doing the work
Content shelf life	Months to build, then ages	Varies	Buried in days	Continuously refreshed
Practice built in	Sometimes	No	No	Yes, every video
Quality and safety review	Yes	Yes	No	Yes, with AI flagging
Shows what people need	Rarely	No	Informally	Requests and votes

Product decisions

AI never approves content. It suggests and flags. A human reviewer always makes the publish decision.
Every AI output is labeled and editable. Summaries, tags, and activities are validated by a person.
Trust is built into the upload. Contributors confirm consent, ownership, confidentiality, and accessibility before a submission can move forward.
Swappable parts. AI and video hosting sit behind adapters, and sign-in can move to single sign-on later without changing the rest of the app.

Questions leaders will ask

Why not just use chat tools or a video site?

Those tools hold conversations and files, but they don't turn a demonstration into a learning experience. There is no review, no practice activity, no skill tagging, and no way to see what is missing. This is built for learning from the start.

What about confidential information?

It is the biggest risk, so it gets the most attention. Contributors confirm consent and confidentiality before submitting. AI scans for things like client names and personal data and points to the exact timestamp. A human reviewer approves before publishing, and anyone can report a video afterward.

Who is going to record these?

Start with the people who already do this informally: the colleague everyone asks for help. The request feature helps too, because people are far more willing to record something when someone asked for it by name.

Can we trust AI-generated summaries?

Every AI output is labeled, editable, and validated by a person. AI never approves, rejects, or removes anything on its own.

What will it cost?

It is best sized after a pilot. The main costs are reviewer time, video hosting (already in place), hosting the app, and AI usage, which scales with the number of videos.

How will we know it works?

By measuring it, using the pilot plan above.

Proposed pilot

Scope: two communities (for example AI at Work and Customer Discovery), 20 to 40 participants, 6 to 8 weeks.
Goal: 10 to 15 published demonstrations and a steady stream of requests.
What we would measure: how many people contribute, how many watch to the end, how many complete the Try activity, how many report using it on real work, how long review takes, and what share of requests get answered.
Decision point: at the end, review the numbers and decide whether to expand, change, or stop.

The scope and targets are starting suggestions to adjust with stakeholders.

Skills shown

Product discovery and framing, prototype-to-product planning, workflow and data modeling, responsible AI design with a human in the loop, privacy and trust design, pilot and measurement planning, and executive storytelling.

This is a prototype in development, described in general terms. Nothing here reflects real user data.