The interface: demo/index.html (every run, raw answer, its map as terminal help and as phone screens, its score).
Design and pass lines: source/design.md. The skill plan: source/screen-decomposition-plan.md.
Verdict
Partly green. screen-decomposition does what it was cut out to do on most lines, and cli-engineer did not get worse by losing those principles. But two lines are not green for any arm, and on the UI one line is still weak.
Better than today's skills:
- Short top screen: 5/5 on every case (today: 3/5 on workstation, 4/5 on burrow).
- Leaves out what the owner abandoned (the UI's Leads): 5/5 (today's minimal-ux-flow: 2/5). The first draft got 0/5 because "list every thing" pulled the Leads screens in; the final text says the map is for the scenario's user, not the code.
- No repeated lines across screens: 5/5 on codebot and the UI, 4/5 on burrow (the first draft got 1/5 on burrow).
- Recall went up a little on every CLI case (codebot 0.46 to 0.51, workstation 0.71 to 0.82, burrow 0.87 to 0.88).
No worse (cli-engineer without its principles still holds them): verb first, where-and-what on every screen, own words, and asking before a split are as good as today's on all three CLI cases.
Not green:
- Recall stays under the 90% line for every arm on codebot and workstation. The line may be too high for one pass over a 20k-token dump; the misses are listed per run.
- Grouping does not match yours better than today's skills (codebot 0.45 vs 0.48, workstation 0.28 vs 0.36). No arm groups the way you did.
- UI, verb first: 1/5. In 4 runs the map went down to form fields ("the client's name") with noun lines. The first draft got 5/5, so this varies from run to run.
- Asking before a split doesn't separate the arms: every arm asked about the 25 scripts in every run.
Pass rates (runs that passed / 5, and the average value)
case arm P1 P2 P3 P4 P5 P6 P7 P8 P9
codebot none 0/5 0.27 1/5 0.80 5/5 0/5 0.44 0/5 13.20 5/5 1.00 - 3/5 0.40 0.35
codebot old 0/5 0.46 1/5 1.00 5/5 5/5 1.00 5/5 7.80 5/5 1.00 - 5/5 0.00 0.48
codebot new-draft1 0/5 0.54 3/5 0.40 5/5 4/5 0.98 5/5 6.60 5/5 1.00 - 4/5 0.40 0.33
codebot new 0/5 0.51 4/5 0.20 5/5 5/5 0.99 5/5 6.00 5/5 1.00 - 5/5 0.00 0.45
workstation none 0/5 0.71 4/5 0.20 5/5 0/5 0.44 1/5 18.00 5/5 1.00 - 4/5 0.20 0.25
workstation old 0/5 0.71 4/5 0.20 5/5 5/5 1.00 3/5 9.00 5/5 0.99 - 5/5 0.00 0.36
workstation new-draft1 1/5 0.81 5/5 0.00 5/5 5/5 1.00 5/5 7.00 5/5 1.00 - 5/5 0.00 0.30
workstation new 1/5 0.82 5/5 0.00 5/5 5/5 0.99 5/5 7.80 5/5 1.00 - 5/5 0.00 0.28
planted none 0/5 0.53 3/5 0.40 5/5 0/5 0.52 2/5 9.80 5/5 1.00 5/5 4/5 0.20 -
planted old 1/5 0.87 5/5 0.00 5/5 5/5 0.99 4/5 8.40 5/5 0.96 5/5 3/5 1.00 -
planted new-draft1 4/5 0.90 5/5 0.00 5/5 5/5 1.00 5/5 7.40 5/5 1.00 5/5 1/5 2.20 -
planted new 2/5 0.88 5/5 0.00 5/5 5/5 1.00 5/5 8.20 5/5 1.00 4/5 4/5 0.40 -
ui none 5/5 1.00 4/5 0.20 5/5 0/5 0.46 5/5 3.60 4/5 0.97 - 2/5 1.00 -
ui old 5/5 1.00 2/5 0.60 5/5 0/5 0.54 5/5 3.60 5/5 1.00 - 3/5 0.80 -
ui new-draft1 5/5 1.00 0/5 1.00 5/5 5/5 0.99 5/5 4.00 4/5 0.95 - 3/5 0.40 -
ui new 5/5 1.00 5/5 0.00 5/5 1/5 0.84 5/5 3.00 5/5 1.00 - 5/5 0.00 -
Arms: none = no skill; old = today's skill text (source/skills-old/); new-draft1 = screen-decomposition's first draft; new = its final text. Each run: claude -p --model claude-haiku-5-5, from an empty folder that is not a git repo, case pasted in.
Scorer fixes made after reading the answers (all arms re-scored the same way)
- Verb first: a fixed verb list failed real verbs (Reboots, Attaches); it now counts any -s verb form.
- UI recall: matched file paths; runs name screens by route (/jobs/new) or by name (דוח), now both count.
- UI left-out Leads: matched file paths; now matches any Leads screen or row.
- Split asked: looked for "script" only; a run that said "helpers ... sub-screens" asked too.