Uploaded October 2026 | Updated November 2026, 3 hours ago
I left six local AI agents working through my Avalonia theme overnight. This is what they got done, what they got wrong, and what I learned about letting models fix UI they can't actually see.
The agents ran on Qwen 27B on my own machine, orchestrated by MiMo in AI Mecca, working from a plan I dictated: screenshot each control in the Theme Lab, check it against the theme rules, fix it and prove the fix with before and after shots. Along the way: turning screenshots into structured data so a model can work out alignment and padding, agents faking their screenshots and breaking the shared build, saving their lessons before the context runs out, and what the same run would have cost on paid models. Then we go through the results side by side, broken on one side and fixed on the other. Part of the series on building a custom Avalonia UI theme from scratch.
━━━━━━━━━━━━━━━━━━━
📌 WHAT YOU'LL LEARN
━━━━━━━━━━━━━━━━━━━
• How a team of local agents can work through every control in a theme overnight
• Why structured data from screenshots matters for models that can't see pixels
• How theme rules stop each control drifting from the shared tokens
• What goes wrong when agents fake their screenshots, and how to catch it
• Why it pays to save the agents' lessons before their context is compacted
• What a run like this would cost in tokens on paid models
• Where the fixes still need a human, from calendar collapses to combo box padding
━━━━━━━━━━━━━━━━━━━
⏱ TIMESTAMPS
━━━━━━━━━━━━━━━━━━━
[0:00] - Recursive self-improvement, or close to it
[0:44] - My personal AI Mecca app
[1:16] - Six local agents running at once
[1:43] - Shushing the agent's voice
[2:02] - First look at the fixed themes
[3:15] - Linking every control to the theme
[3:52] - It fixed itself
[4:29] - The plan behind the overnight run
[5:00] - The structured data breakthrough
[6:21] - Dictating the rules
[7:05] - Watching the agents work
[7:41] - Fake screenshots and a broken build
[8:20] - Saving the lessons before compaction
[9:52] - Why I'm so excited about this
[10:34] - What it would have cost in tokens
[11:28] - Why models don't see pixels
[12:40] - Wrap-up, before and after
[14:36] - Side by side, fixed against broken
[16:00] - Calendar collapse and padding
[17:05] - Checkboxes and SVG ticks
[18:33] - Combo boxes and the rest
[19:45] - Training models on the results
[20:47] - Final thoughts
━━━━━━━━━━━━━━━━━━━
🔗 LINKS & RESOURCES
━━━━━━━━━━━━━━━━━━━
▶ Full Avalonia Themes Playlist: youtube.com/playlist?list=PLrW43fNmjaQVE2vDBeWD8QYIm06w6bKyO
▶ Source Code (AvaloniaThemeLab): github.com/angelsix/youtube/tree/develop/AvaloniaThemeLab
▶ GitHub: github.com/angelsix
▶ Website: angelsix.com
━━━━━━━━━━━━━━━━━━━
📚 SERIES - Avalonia UI Theming
━━━━━━━━━━━━━━━━━━━
▶ Full series: youtube.com/playlist?list=PLrW43fNmjaQVE2vDBeWD8QYIm06w6bKyO
19 - Calendar Control From Scratch - youtube.com/watch?v=EZE8MFbcoAE
20 - Build a Light vs Dark Preview Control - youtube.com/watch?v=tYTMNrN-3rw
21 - Hover Cropping, Corners and Contrast - youtube.com/watch?v=Oih-a3t687E
22 - Code Reviewing Is Boring - youtube.com/watch?v=yiZcrImMsXg
23 - This Video - An overnight AI agent run on the theme
I left six local AI agents working through my Avalonia theme overnight. This is what they got done, what they got wrong, and what I learned about letting models fix UI they can't actually see.
The agents ran on Qwen 27B on my own machine, orchestrated by MiMo in AI Mecca, working from a plan I dictated: screenshot each control in the Theme Lab, check it against the theme rules, fix it and prove the fix with before and after shots. Along the way: turning screenshots into structured data so a model can work out alignment and padding, agents faking their screenshots and breaking the shared build, saving their lessons before the context runs out, and what the same run would have cost on paid models. Then we go through the results side by side, broken on one side and fixed on the other. Part of the series on building a custom Avalonia UI theme from scratch.
━━━━━━━━━━━━━━━━━━━
📌 WHAT YOU'LL LEARN
━━━━━━━━━━━━━━━━━━━
• How a team of local agents can work through every control in a theme overnight
• Why structured data from screenshots matters for models that can't see pixels
• How theme rules stop each control drifting from the shared tokens
• What goes wrong when agents fake their screenshots, and how to catch it
• Why it pays to save the agents' lessons before their context is compacted
• What a run like this would cost in tokens on paid models
• Where the fixes still need a human, from calendar collapses to combo box padding
━━━━━━━━━━━━━━━━━━━
⏱ TIMESTAMPS
━━━━━━━━━━━━━━━━━━━
[0:00] - Recursive self-improvement, or close to it
[0:44] - My personal AI Mecca app
[1:16] - Six local agents running at once
[1:43] - Shushing the agent's voice
[2:02] - First look at the fixed themes
[3:15] - Linking every control to the theme
[3:52] - It fixed itself
[4:29] - The plan behind the overnight run
[5:00] - The structured data breakthrough
[6:21] - Dictating the rules
[7:05] - Watching the agents work
[7:41] - Fake screenshots and a broken build
[8:20] - Saving the lessons before compaction
[9:52] - Why I'm so excited about this
[10:34] - What it would have cost in tokens
[11:28] - Why models don't see pixels
[12:40] - Wrap-up, before and after
[14:36] - Side by side, fixed against broken
[16:00] - Calendar collapse and padding
[17:05] - Checkboxes and SVG ticks
[18:33] - Combo boxes and the rest
[19:45] - Training models on the results
[20:47] - Final thoughts
━━━━━━━━━━━━━━━━━━━
🔗 LINKS & RESOURCES
━━━━━━━━━━━━━━━━━━━
▶ Full Avalonia Themes Playlist: youtube.com/playlist?list=PLrW43fNmjaQVE2vDBeWD8QYIm06w6bKyO
▶ Source Code (AvaloniaThemeLab): github.com/angelsix/youtube/tree/develop/AvaloniaThemeLab
▶ GitHub: github.com/angelsix
▶ Website: angelsix.com
━━━━━━━━━━━━━━━━━━━
📚 SERIES - Avalonia UI Theming
━━━━━━━━━━━━━━━━━━━
▶ Full series: youtube.com/playlist?list=PLrW43fNmjaQVE2vDBeWD8QYIm06w6bKyO
19 - Calendar Control From Scratch - youtube.com/watch?v=EZE8MFbcoAE
20 - Build a Light vs Dark Preview Control - youtube.com/watch?v=tYTMNrN-3rw
21 - Hover Cropping, Corners and Contrast - youtube.com/watch?v=Oih-a3t687E
22 - Code Reviewing Is Boring - youtube.com/watch?v=yiZcrImMsXg
23 - This Video - An overnight AI agent run on the theme
![One Token System for the Entire Avalonia Theme | Dev + AI
Hundreds of random hardcoded values scattered across every control in the theme. Margins,
thicknesses, font sizes — all over the place, inconsistent, unmaintainable. In this episode
we do the deduplication pass: collapsing everything down to a minimal, precise set of design
tokens that defines the entire theme system.
AI does the grunt work again — but the interesting part is what happens when it goes off-brief.
Real-time corrections, naming convention debates, a new Scaled markup extension that solves
proportional scaling for edge-case values, and a moment where the AI adds things I explicitly
told it not to. This is what collaborative AI development actually looks like in practice.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📌 WHAT YOULL LEARN
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
• How to use AI to find and replace hundreds of scattered hardcoded values across a theme
• Adding new token properties to the engine: font family, font sizes, control height, icon
• What the Adorner Layer is and how to correctly map its thickness values to tokens
• How to establish naming conventions for thickness and spacing tokens — with real examples
• When NOT to add a value to the token set — the token bloat problem and how to avoid it
• Styling AutoCompleteBox: placeholder text, border brushes, padding, drop-down height
• The Scaled markup extension — what it is, why its needed, and how to implement it right
• How to handle negative margins using Edges + a negative multiplier
• How to correct AI when it goes off-brief — real examples of steering it back in real-time
• The ultimate goal: swap any compliant theme without a single line of application code
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⏱ TIMESTAMPS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[0:00] – Introduction — new theme engine recap, goal of the deduplication pass
[0:25] – Using AI to find and replace legacy dynamic resources
[0:48] – New token properties added: font family, font sizes, control height, icon
[1:04] – Button control reviewed — already clean, moving on
[1:31] – What the Adorner Layer is and why it draws on its own visual layer
[3:08] – Prompting AI to merge AdornerLayer thickness values into the theme engine
[3:56] – Moving to AutoCompleteBox — placeholder, borders, padding
[8:49] – AIs first response on AdornerLayer thickness mapping
[11:08] – Decision: add dedicated thickness tokens to the engine
[11:59] – AI identifies further inline resources across other controls
[13:41] – Asking AI for thickness token naming conventions
[13:52] – AIs full token proposal reviewed and critiqued
[14:26] – The token bloat problem — when NOT to add a new token
[18:11] – Policy decision: how to snap stray values to existing tokens
[20:08] – AutoCompleteBox: border brush and accent brush tokens
[21:45] – MaxDropDownHeight and margin replacement
[23:27] – ⚠️ AI adds unwanted spacing values — correcting it in real-time
[24:17] – Introducing the Scaled markup extension for proportional custom values
[26:09] – Problem: Scaled extension hardcoded to ThemeContext — feedback given to AI
[28:17] – Further AI instructions on the Scaled extension + remaining issues
[32:59] – Bug fix discovered in Avalonia source code for AutoCompleteBox
[33:42] – Negative margins fixed with Edges + negative multiplier
[34:07] – Git diff review: AdornerLayer and AutoCompleteBox fully cleaned up
[36:45] – The ultimate goal: seamless theme swapping without any application code changes
[38:00] – Wrap up & next steps
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔗 LINKS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
▶ Part 1 – Setup: https://www.youtube.com/watch?v=rUKlQnjU0tc
▶ Part 2 – Buttons: https://www.youtube.com/watch?v=oKjk6pyyARI
▶ Part 3 – Button Spinner Anatomy: https://www.youtube.com/watch?v=WMaTbXkvm9g
▶ Part 4 – Designing a Theme System: https://www.youtube.com/watch?v=I5halHoNfH0
▶ Part 5 – Checkbox Anatomy: https://www.youtube.com/watch?v=uIimnwjITT0
▶ Part 6 – Checkbox Styling: https://www.youtube.com/watch?v=uv1bNrfCjtY
▶ Part 7 – Theme Engine From Scratch: https://www.youtube.com/watch?v=NXz8hwITg5U
▶ Part 8 – Design Token System: https://www.youtube.com/watch?v=QMVdxRZqaU8
▶ Part 9 – Stripping Legacy Code: https://www.youtube.com/watch?v=5-p_fnOi0As
📺 Full Series Playlist: https://www.youtube.com/playlist?list=PLrW43fNmjaQVE2vDBeWD8QYIm06w6bKyO
💾 Source Code: https://github.com/angelsix/youtube/tree/develop/AvaloniaThemeLab
🌐 AngelSix: https://www.angelsix.com
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📺 SERIES: Avalonia UI Theming – Hand Coded From Scratch
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
#avalonia #avaloniaui #xaml #csharp #dotnet #aiagent #aitools #softwaretutorial #coding One Token System for the Entire Avalonia Theme | Dev + AI](https://i.ytimg.com/vi/JEcD1XXIiA8/mqdefault.jpg)









