Michael Limberger
Need me? Email mike@limberger.ca
Skiff
HTML · resume page
One of the three model tests
This is the web-shaped task.
Each model had to read a fixture resume in markdown, then write a
professional resume.html with inline CSS, real facts,
and no CDN. I scored capability with the checklist harness, then
gave a separate vibe grade for how the page looked.
North’s first pass had double-bullet artifacting, so I gave it another try. The times below are the retry. Harness stayed at 100%. Vibe moved from C- to B.
The task
Read the fixture resume markdown. Write
resume.html with write_file. Keep CSS inline. Use the
real facts from the fixture. Stay inside the workspace.
Harness
Checklist scorer. Most models hit 100% here, so the grade leans on hygiene, turns, and wall time.
Vibe
My subjective look at color, interest, and polish. Separate from the harness letter.
Local MLX
| Model | Time | Pass | Vibe |
|---|---|---|---|
| ornith | 81.3s | 100% | A- |
| qwen36 | 86.8s | 100% | B- |
| kat | 116.3s | 100% | A |
| ministral | 118.3s | 100% | B |
| north | 197.7s | 100% | B |
| Devstral | 176.2s | 100% | B |
| bonsai | 311.2s | 100% | B+ |
kat still leads vibe at A. ornith is right behind at A-, and it was the fastest clean resume. bonsai finished, but it took a long time.
Ollama Cloud
| Model | Time | Pass | Vibe |
|---|---|---|---|
| glm | 14.7s | 100% | B- |
| flash | 15.2s | 100% | B+ |
Both models sat around 15 seconds. flash looked better. glm was a hair cleaner on the tool path in the combined HTML plus Bash round.
Sample pages
A few of the generated resumes. These open as plain HTML from this site.
kat
A vibe
ornith
A- vibe · fastest local
flash
B+ vibe · cloud
glm
B- vibe · cloud
north
B vibe after retry
bonsai
B+ vibe · slowest
All boards together live on the model tests page.