Initiative case · AI enablement · Designing with AIzvoove · staffing software · 2025 – 2026
My working rules became the team’s AI infrastructure.
As the only designer for several teams, I built myself workflows and skills to keep up. On my own initiative this became zvoove Skills: the shared AI library of Product, Design and Engineering.
- Role in the project
- InitiatorOwn initiative, no brief
- Setting
- AI Steering (Value Builders)Founding member
- Responsibility
- Skills and foundationzvoove Brain, design and PM skills, knowledge base and design system connected, guardrails
- Status
- and 2 more states
Live+2
As of September 2026
Live
Since March 2026: 19 skills at launch, 34 across three departments in August 2026
In use
Dark Factory with the developers since August 2026, the pipeline belongs to Engineering
Open
Measured internally by tickets done
The path
01
Solve my own problem first
write the way I work down instead of repeating it.
Starting point
What slowed things down
- Design was brought in too late
- AI did not fit the system
- Context got lost
The decision
Instead of explaining the same things every day, I wrote down how I work. At first just for me. Then it turned out others could work with it just as well.
02
Connect knowledge instead of copying it
the knowledge base via MCP, improved together with its team.
MCP
- Which fields does a timesheet export for customers need?
knowledge-base · search3 results- 1Timesheet · data modelEntity, 14 fields
- 2Export · CSV and PDFGuideline
- 3Persona · dispatcherPersona
A customer export needs 9 of the 14 fields: employee, assignment, date, start, end, breaks, hours, surcharges and approval. Internal cost centres stay out.
1Timesheet · data model2Export · CSV and PDF3Persona · dispatcher
Write a message …Rebuilt: an agent queries the Knowledge Base over MCP instead of someone pasting context into every session.
03
Product managers build for themselves
PM skills and an AI-ready design system.
Product
Until the first customer feedback BeforeFeedback after the build
- 1IdeaThe PM's rough thought.
- 2TicketA ticket without a screen.
- 3Waiting for designOne designer for several teams: the ticket waits.
- 4DraftDesign drafts, the PM aligns.
- 5BuildEngineering builds. Only then do customers see anything.
- Customer feedbackOnly after the build: changes now mean rework.
With the skillsFeedback before the first ticket
no detourNo detour to the feedbackTicket, waiting for design, draft and build now come after the first customer feedback, no longer before it.- 1IdeaThe PM's rough thought.
- 2Spec → prototypePM skill and prototype skill: the thought becomes a spec, and from it, automatically, a near-live prototype in UnityUI.
- Customer feedbackOn a near-live prototype, before a ticket exists.
- No detour to the feedbackTicket, waiting for design, draft and build now come after the first customer feedback, no longer before it.
with skillsWaitingCustomer feedbackPM skillPrototype skill04
A chain instead of single prompts
from idea to handover.
chain
The chainfive stations 01Spec
A sentence, a Figma link or meeting notes become a buildable spec: persona, problem, user stories, acceptance criteria, edge cases.
- Persona
- Problem
- User stories
- Acceptance criteria
- Edge cases
02Design refinement
Every ticket with a user-facing surface gets its design task. It covers flow, error behaviour, states, permissions, copy and translation and stays in sync with the ticket.
- Flow
- Error behaviour
- States
- Permissions
- Copy
- Translation
03Figma to UnityUI
Spacing, typography, colours and layout come from Figma and are translated into UnityUI code, checked against the design tokens.
- Spacing
- Typography
- Colour
- Layout
checked against the design tokens
04Quality gate
Every component is checked from six angles: code quality, design system compliance, accessibility, security, performance, completeness, with severities from P0 to P3.
- Code quality
- Design system fidelity
- Accessibility
- Security
- Performance
- Completeness
Severities P0 to P3
05Handover
A ten-part package with context, architecture decisions, test instructions and a quick start for reviewers.
- Context
- Architecture decisions
- Test guide
- Reviewer quick start
ten-part package
05
No effect without a yes
every action waits for approval.
approval
Claude Code - /pm-ticket “Export timesheets as CSV”
- Spec: persona, problem, user stories, acceptance criteria, edge cases
- Design task: flow, error behaviour, states, permissions, copy
- Create the ticket and comment in the team?
- Ticket created and commented in the team.
- Cancelled. Nothing created.
06
Let go to scale
a shared structure with reviews and owners.
team
From tool to platform - 01For meWorkflows and skills to cope as the only designer.
- 02ConnectedKnowledge base via MCP, access improved with its team.
- 03For product managersPM skills and the AI-ready design system.
- 04zvoove BrainEnablement for the whole company.
- 05zvoove SkillsContinued within the AI Steering (Value Builders), with reviews and owners.
07
Supply the material, not the pipeline
skills and patterns for the Dark Factory.
agents
Dark Factory 01Export timesheets as CSVcreated with the PM ticket skill02Pipelineagents implement · Engineering03Quality gate04Codefor review
ZVS-1482
Export timesheets as CSV
darkfactoryon
Pipeline
Quality gateP0–P3
- Code quality
- Design system compliance
- Accessibility
- Security
- Performance
- Completeness
Code
Pull request
Export timesheets as CSV
Checks passed
waits for review from Engineering
My part
From zvoove Skills. Engineering owns the pipeline; my part is the material it draws on.
- PM ticket skillmine, carried on with the team
- Design refinement skillextended for the Dark Factory
- UnityUI pattern librarybeing expanded
Level 5 of 5agents implement tickets on their own
08
From 19 to 34 skills
Today within the AI Steering (Value Builders), with reviews and six owners.
today
The libraryRepository snapshots - 13 Mar 202619 skillsfour plugin packages, all by me
- August 202634 skillsProduct, Design and Engineering, six owners
12 by me22 by five colleagues
From the library, a selection
- design-refinementDesign
- unity-ui-qaDesign
- unity-ui-pattern-genDesign
- unity-ui-storybookDesign
- product-specProduct
- discovery-to-prototypeProduct
- zvoove-wording-checkerProduct
- user-insightsProduct
- dev-handoverEngineering
- jira-to-buildEngineering
My working rules became the team’s AI infrastructure.
One designer for many teams only works with a system
At zvoove Germany I was the only designer, responsible for several teams and products at once. Every team needed design, and I could not be everywhere at the same time.
So I started taking work off my own plate. I had been building my own workflows for AI tools since September 2024; now I wrote workflows and skills that held research, personas, design system rules and constraints, so I no longer had to explain them in every session.
Design was brought in too late
Questions about states, errors, permissions and copy only surfaced during implementation.
AI did not fit the system
Output from AI tools ignored UnityUI: hard-coded values, wrong components, crashes.
Context got lost
Personas, domain terms like the German temporary-staffing law and product wording had to be explained again in every session.
Design did not scale
Few designers, many teams: every question waited for a person.
No shared rules
Everyone worked with AI differently, and nobody owned it.
On top of that, product development as a whole was slow. What slowed me down as a lone designer would, over time, slow everyone down.
Connect knowledge instead of copying it
In parallel, zvoove built a knowledge base from the database records about our products. It was meant to let customer support pull up any data through an AI chat and answer tickets and calls more easily.
I connected it to my workflows through its MCP access. To make that just as easy for others, I worked closely with the team that runs the knowledge base and improved the MCP access together with them.
Knowledge base
Product data from the database, reachable via MCP.
My workflows
Connected via the MCP access, improved together with the team.
Customer support
AI chat for tickets and calls.
zAIn
From proof of concept to the AI layer in the product.
The same knowledge base led to something else. The PM lead, the tech lead and I had the same idea: turn it into an AI assistant for the product. We started an initiative group, built a proof of concept and pushed it internally. It became zAIn and later the whole AI layer in the product.
Learning
A skill is only as good as the knowledge it is connected to. Improving the access was worth more than any additional prompt.
Product managers build prototypes that feel real
Once my own workflows held up, I started writing skills for product managers: from a rough thought to a structured spec, from a ticket to a design task.
At the same time I brought in the design system. UnityUI is an existing system that I did not create. Separately from the skills, I made it AI-ready from autumn 2025, first with Figma Make, later together with the frontend lead. Since then it works in Figma Make, Claude Design, Claude Code, Codex and Lovable.
Strand 1
PM skills
Specs, tickets and design tasks in the same structure, with personas and domain terms from the shared foundation.
Strand 2
AI-ready design system
UnityUI with tokens and components that AI tools apply correctly.
Result
Near-live prototypes
Product managers use them to validate on their own, run discovery and back up tickets and requirements. We work together in Claude Design or Figma Make until I take over the detailing and the states.
BeforeFeedback after the build
- 1IdeaThe PM's rough thought.
- 2TicketA ticket without a screen.
- 3Waiting for designOne designer for several teams: the ticket waits.
- 4DraftDesign drafts, the PM aligns.
- 5BuildEngineering builds. Only then do customers see anything.
- Customer feedbackOnly after the build: changes now mean rework.
With the skillsFeedback before the first ticket
- 1IdeaThe PM's rough thought.
- 2Spec → prototypePM skill and prototype skill: the thought becomes a spec, and from it, automatically, a near-live prototype in UnityUI.
- Customer feedbackOn a near-live prototype, before a ticket exists.
- No detour to the feedbackTicket, waiting for design, draft and build now come after the first customer feedback, no longer before it.
Before, the first customer feedback came only after the build: the idea became a ticket without a screen, waited for design, went through draft and alignment, and then into the build. With the skills, product managers get feedback on a near-live prototype, before the first ticket.
AI Playbook 2.0 Lite
For my PM colleagues I wrote the AI Playbook 2.0 Lite: how to build with Claude, Claude Design and Claude Code, how to plug in skills from the zvoove Skills plugin and what to use for what, with weekly learnings and shared ownership.
From idea to handover, step by step
Single prompts help one person. An organisation needs a chain in which each step picks up the previous result, checks it and passes it on. On 13 March 2026 it sat in the repository as 19 skills in four plugin packages, all written by me.
01Spec
A sentence, a Figma link or meeting notes become a buildable spec: persona, problem, user stories, acceptance criteria, edge cases.
- Persona
- Problem
- User stories
- Acceptance criteria
- Edge cases
02Design refinement
Every ticket with a user-facing surface gets its design task. It covers flow, error behaviour, states, permissions, copy and translation and stays in sync with the ticket.
- Flow
- Error behaviour
- States
- Permissions
- Copy
- Translation
03Figma to UnityUI
Spacing, typography, colours and layout come from Figma and are translated into UnityUI code, checked against the design tokens.
- Spacing
- Typography
- Colour
- Layout
checked against the design tokens
04Quality gate
Every component is checked from six angles: code quality, design system compliance, accessibility, security, performance, completeness, with severities from P0 to P3.
- Code quality
- Design system fidelity
- Accessibility
- Security
- Performance
- Completeness
Severities P0 to P3
05Handover
A ten-part package with context, architecture decisions, test instructions and a quick start for reviewers.
- Context
- Architecture decisions
- Test guide
- Reviewer quick start
ten-part package
No effect without a yes
AI that creates tickets or writes comments needs clear limits. So the same rule applies to every skill: a skill starts on its own when the request fits, but every action with a real effect waits for explicit approval.
- /pm-ticket “Export timesheets as CSV”
- Spec: persona, problem, user stories, acceptance criteria, edge cases
- Design task: flow, error behaviour, states, permissions, copy
- Create the ticket and comment in the team?
- Ticket created and commented in the team.
- Cancelled. Nothing created.
Approval before effect
Tickets, comments and messages only happen after a yes.
One shared foundation
Personas, glossary, product wording and AI guardrails live in one central place. Every skill refers to them instead of copying them.
Reviews per department
Changes go through pull requests. Design, product and engineering each review their own area.
Automatic checks
Every skill is validated automatically before it is merged.
On top of that comes the AI Playbook I wrote: general rules for working with AI, with prompting help, direction, mission and vision, and the call to label AI-built work. It is used across the group. In August 2026 the skill catalogue followed, which makes the skills findable for everyone.
My tool became infrastructure
My workflows became zvoove Brain: the idea of tying the company’s knowledge to roles, workflows and processes, so everyone has what they need at hand and nobody has to search. Enablement for the whole company, started without a mandate.
- 01For meWorkflows and skills to cope as the only designer.
- 02ConnectedKnowledge base via MCP, access improved with its team.
- 03For product managersPM skills and the AI-ready design system.
- 04zvoove BrainEnablement for the whole company.
- 05zvoove SkillsContinued within the AI Steering (Value Builders), with reviews and owners.
The VP of Engineering founded the AI Steering (Value Builders) to steer AI topics and enablement at zvoove Germany. I am a founding member. That is where zvoove Brain became zvoove Skills. I am responsible for the skills, including all design skills.
- 13 Mar 202619 skillsfour plugin packages, all by me
- August 202634 skillsProduct, Design and Engineering, six owners
12 by me22 by five colleagues
The most important result is not the number of skills but the owner entries: colleagues from other teams now write and maintain skills of their own. The library is distributed as a plugin marketplace for Claude Code and Cursor.
Automation is only as good as the system it reads
Dark Factory is the level at which agents implement tickets on their own, the fifth level in Dan Shapiro’s public maturity model for AI-assisted coding. At zvoove it started as a pilot in August 2026 and is now in use by the developers. Engineering owns the pipeline.
My part is the material it draws on. A ticket created with my PM ticket skill can be handed to the Dark Factory with “darkfactory on”. To improve its results, I started extending the design refinement skill for it and building out the UnityUI pattern library.
01Export timesheets as CSVcreated with the PM ticket skill02Pipelineagents implement · Engineering03Quality gate04Codefor review
ZVS-1482
Export timesheets as CSV
darkfactoryon
Pipeline
Quality gateP0–P3
- Code quality
- Design system compliance
- Accessibility
- Security
- Performance
- Completeness
Code
Pull request
Export timesheets as CSV
Checks passed
waits for review from Engineering
My part
From zvoove Skills. Engineering owns the pipeline; my part is the material it draws on.
- PM ticket skillmine, carried on with the team
- Design refinement skillextended for the Dark Factory
- UnityUI pattern librarybeing expanded
Level 5 of 5agents implement tickets on their own
Learning
Automation moves work rather than removing it. The fewer people stand between ticket and code, the more precise the rules, patterns and skills have to be.
What came of it
Connect knowledge
Knowledge base connected via MCP, shared foundation
Stop explaining context in every session
Enable product managers
PM skills and AI-ready UnityUI in five tools
Prototypes without a detour through design
Scale
34 skills, six owners (August 2026)
from one person to the organisation
Anchor
AI Steering (Value Builders), reviews per department
Responsibility beyond me
Automate
Dark Factory in use by the developers
Implement tickets without manual work
What is still openopen
internally by tickets done
Measure the effect
Conclusion
I did not introduce AI. I wrote down how we work and connected it so that AI can work by it.
What I learned and would do differently today
Caveats
What is still open
It is measured internally by tickets done. What is documented is scope, structure and who contributes skills.
From the handover on, it is team work
Since the library moved into the AI Steering group, many people grow it: 22 of the 34 skills come from others.
The pipeline is not mine
Engineering owns the Dark Factory. My share is the material it draws on.
I did not create UnityUI
I made the existing system AI-ready, first on my own, then with the frontend lead.
zAIn was a joint initiative
The PM lead, the tech lead and I had the idea for the assistant together. I started the skills on my own.
As of September 2026
All information reflects that point in time.
Learnings
Where I get stuck, the organisation soon will
Carrying many teams on your own, you see early where the system jams. What slowed me down would soon have slowed everyone. I solved it before it got there.
Knowledge beats prompts
The biggest leaps came from better access to knowledge, not from better wording.
Guardrails create trust
Because nothing happens without a yes, tasks can be handed to AI without giving up control.
Letting go is scaling
The library has grown since it stopped belonging only to me.
What I would do differently today
Measure from day one
Before starting, I would record a baseline: how long it takes from ticket to handover and how often design questions come too late. Without it, the effect stays a good story instead of evidence.
Look for a frame earlier
I built on my own for a long time. Reviews and owners came with the AI Steering group. I could have looked for that frame earlier instead of adding it afterwards.
Plan for maintenance
The AI Playbook is used across the group but no longer maintained. Today I would decide from the start who maintains each artefact, the way the skills now have their owners.
Start with a few skills and real teams
Fewer skills, tried earlier with individual teams, would have shown sooner which ones hold up.
I built it to understand how something like this works. I brought it into the company because a system is only worth something once others work with it.
fin (for now)