Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Provides sem semantic-diff detection, install-on-first-use, and fallback patterns. Use when building skills that consume git diff output.
.claude/skills/athola-sem-integration/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -2% | 0% |
Foundation patterns for using sem semantic diffs in night-market skills.
Consult this skill when building or modifying skills that consume git diff output. It provides the detection, installation, and fallback patterns.
sem diff yourself)Check sem availability once per session:
bash# Detection (cache per session) _sem_check() { local cache="${CLAUDE_CODE_TMPDIR:-/tmp}/sem-available" if [ -f "$cache" ]; then cat "$cache" return fi if command -v sem &>/dev/null; then echo "1" | tee "$cache" else echo "0" | tee "$cache" fi }
When _sem_check returns 0, offer installation. See modules/detection.md for install-on-first-use logic and platform-specific commands.
Primary path (sem available):
bashsem diff --format json <baseline>
Fallback path (sem unavailable):
bashgit diff --name-only --diff-filter=A <baseline> git diff --name-only --diff-filter=M <baseline> git diff --name-only --diff-filter=D <baseline> git diff --name-only --diff-filter=R <baseline>
See modules/fallback.md for output normalization that produces the same entity schema from both paths.
Primary path (sem available):
bashsem impact --json <file-or-entity>
Fallback path (sem unavailable): use rg/grep to trace callers by filename. See modules/fallback.md.
_sem_check result cached per session at${CLAUDE_CODE_TMPDIR:-/tmp}/sem-available; detection command runs at most once per session regardless of how many skills invoke it
path used automatically (git diff --name-only --diff-filter= variants); no silent failure
JSON output per modules/fallback.md; consumers downstream see the same structure from both paths
dependencies: [leyline:sem-integration] in frontmatter rather than reimplementing the detection pattern
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 9,119 | 8,136 | -11% | 1 | 1 | 0% | 1,748 | 2,395 | +37% | 0 | 0 | — |
case-02 | fail→pass | 17,371 | 10,844 | -38% | 1 | 1 | 0% | 3,568 | 2,719 | -24% | 0 | 0 | — |
case-03 | fail→pass | 15,746 | 14,973 | -5% | 1 | 1 | 0% | 3,037 | 3,728 | +23% | 0 | 0 | — |
case-04 | fail→pass | 8,076 | 3,656 | -55% | 1 | 1 | 0% | 1,337 | 1,145 | -14% | 0 | 0 | — |
case-05 | pass→fail | 11,777 | 10,087 | -14% | 1 | 1 | 0% | 2,075 | 2,423 | +17% | 0 | 0 | — |
case-06 | pass→pass | 11,618 | 8,020 | -31% | 1 | 1 | 0% | 1,959 | 1,968 | +0% | 0 | 0 | — |
case-07 | fail→pass | 8,056 | 4,480 | -44% | 1 | 1 | 0% | 1,497 | 1,460 | -2% | 0 | 0 | — |
case-08 | fail→pass | 17,293 | 3,172 | -82% | 1 | 1 | 0% | 1,034 | 1,108 | +7% | 0 | 0 | — |
case-09 | pass→pass | 4,452 | 3,011 | -32% | 1 | 1 | 0% | 807 | 1,117 | +38% | 0 | 0 | — |
case-10 | pass→pass | 6,301 | 2,802 | -56% | 1 | 1 | 0% | 1,032 | 1,005 | -3% | 0 | 0 | — |
case-11 | pass→pass | 6,836 | 2,016 | -71% | 1 | 1 | 0% | 1,230 | 938 | -24% | 0 | 0 | — |
case-12 | fail→pass | 6,697 | 2,650 | -60% | 1 | 1 | 0% | 1,168 | 1,056 | -10% | 0 | 0 | — |
case-13 | fail→pass | 11,598 | 1,842 | -84% | 1 | 1 | 0% | 1,953 | 890 | -54% | 0 | 0 | — |
case-14 | pass→pass | 12,078 | 5,964 | -51% | 1 | 1 | 0% | 1,970 | 1,461 | -26% | 0 | 0 | — |
case-15 | fail→pass | 6,521 | 1,644 | -75% | 1 | 1 | 0% | 997 | 878 | -12% | 0 | 0 | — |
case-16 | pass→pass | 9,725 | 2,701 | -72% | 1 | 1 | 0% | 1,739 | 1,091 | -37% | 0 | 0 | — |
case-17 | pass→pass | 17,252 | 8,167 | -53% | 1 | 1 | 0% | 3,078 | 2,144 | -30% | 0 | 0 | — |
case-18 | pass→pass | 8,658 | 1,420 | -84% | 1 | 1 | 0% | 1,417 | 821 | -42% | 0 | 0 | — |
case-19 | fail→pass | 8,535 | 2,224 | -74% | 1 | 1 | 0% | 1,658 | 993 | -40% | 0 | 0 | — |
case-20 | pass→pass | 8,845 | 1,971 | -78% | 1 | 1 | 0% | 1,457 | 957 | -34% | 0 | 0 | — |
case-21 | pass→pass | 17,184 | 1,595 | -91% | 1 | 1 | 0% | 2,970 | 857 | -71% | 0 | 0 | — |
case-22 | fail→pass | 12,475 | 9,297 | -25% | 1 | 1 | 0% | 1,980 | 2,185 | +10% | 0 | 0 | — |
case-23 | fail→pass | 11,593 | 5,299 | -54% | 1 | 1 | 0% | 1,753 | 1,495 | -15% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +48 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.