Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use the Agentlas browser hardpoint for browser-required work.
.claude/skills/agentlas-ai-hep-browser/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -82% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -32% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -52% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -45% | 0% |
Update fallback: 자동 업데이트가 안 되면 hephaestus update를 한 번 실행하세요. 업데이트하지 않아도 현재 버전 명령은 그대로 동작합니다.
Raw arguments: everything the user typed after /skill:hep-browser.
Codex plugins cannot register slash commands, so this custom prompt is the explicit entrypoint: /prompts:hep-browser.
Use this when the task needs a real browser, page rendering, login-visible state, click/form behavior, or JS-heavy evidence. Prefer the Agentlas browser hardpoint first.
Behavior:
/prompts:hep-browser <url> reads a rendered browser snapshot./prompts:hep-browser <url> "<instruction>" or/prompts:hep-browser <url> --act "<instruction>" performs browser automation through browser.agent_cli, then captures a snapshot.
by default so logged-in Desktop/browser state is reused.
URLs. For example, use https://mail.google.com/ instead of https://mail.google.com/mail/u/0/#inbox unless that exact deep route is required. Add --raw-url to preserve the exact URL.
--click @e1 or --click-text "Compose" for LLM-free browserprimitives after the host has selected a ref/text target.
--wait-ms 2000 when a dynamic app opens dialogs after a short delay.--read to force snapshot-only mode when extra text is context ratherthan an action.
bashRUNNER="" for c in "$HOME/.agentlas/runtime/current/bin/hephaestus" ./bin/hephaestus; do [ -x "$c" ] && RUNNER="$c" && break done [ -n "$RUNNER" ] || { echo "Hephaestus runtime not found. Run the installer first." >&2; exit 1; } "$RUNNER" hep-browser "$ARGUMENTS"
Report whether browser.agent_cli mounted, whether setup is needed, and the receipt id. For automation, also report the action, whether CDP/profile flags were used, and the final snapshot status. If setup is needed, show hep-browser --setup and hep-browser --check.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 11,375 | 11,663 | +3% | 1 | 1 | 0% | 764 | 1,074 | +41% | 0 | 0 | — |
case-02 | fail→fail | 10,305 | 11,000 | +7% | 1 | 1 | 0% | 1,474 | 1,243 | -16% | 0 | 0 | — |
case-03 | pass→pass | 8,630 | 9,251 | +7% | 1 | 1 | 0% | 1,529 | 2,018 | +32% | 0 | 0 | — |
case-04 | pass→pass | 18,947 | 19,813 | +5% | 1 | 1 | 0% | 3,284 | 4,272 | +30% | 0 | 0 | — |
case-05 | pass→pass | 8,413 | 6,280 | -25% | 1 | 1 | 0% | 1,249 | 1,504 | +20% | 0 | 0 | — |
case-06 | fail→fail | 14,689 | 15,014 | +2% | 1 | 1 | 0% | 1,076 | 1,378 | +28% | 0 | 0 | — |
case-07 | fail→pass | 10,389 | 4,870 | -53% | 1 | 1 | 0% | 1,435 | 1,321 | -8% | 0 | 0 | — |
case-08 | fail→pass | 35,006 | 3,534 | -90% | 1 | 1 | 0% | 5,109 | 918 | -82% | 0 | 0 | — |
case-09 | fail→pass | 9,583 | 3,302 | -66% | 1 | 1 | 0% | 1,495 | 1,015 | -32% | 0 | 0 | — |
case-10 | fail→pass | 14,290 | 3,749 | -74% | 1 | 1 | 0% | 2,104 | 1,015 | -52% | 0 | 0 | — |
case-11 | pass→pass | 10,573 | 4,883 | -54% | 1 | 1 | 0% | 1,492 | 1,216 | -18% | 0 | 0 | — |
case-12 | fail→pass | 11,602 | 4,426 | -62% | 1 | 1 | 0% | 1,715 | 939 | -45% | 0 | 0 | — |
case-13 | fail→pass | 14,700 | 2,770 | -81% | 1 | 1 | 0% | 2,333 | 869 | -63% | 0 | 0 | — |
case-14 | fail→pass | 15,273 | 5,608 | -63% | 1 | 1 | 0% | 2,052 | 1,426 | -31% | 0 | 0 | — |
case-15 | fail→fail | 10,778 | 3,748 | -65% | 1 | 1 | 0% | 1,364 | 1,028 | -25% | 0 | 0 | — |
case-16 | fail→fail | 15,263 | 6,033 | -60% | 1 | 1 | 0% | 2,317 | 1,338 | -42% | 0 | 0 | — |
case-17 | fail→pass | 11,495 | 3,801 | -67% | 1 | 1 | 0% | 1,535 | 834 | -46% | 0 | 0 | — |
case-18 | fail→pass | 13,213 | 4,578 | -65% | 1 | 1 | 0% | 1,880 | 1,063 | -43% | 0 | 0 | — |
case-19 | fail→pass | 19,890 | 2,729 | -86% | 1 | 1 | 0% | 1,635 | 860 | -47% | 0 | 0 | — |
case-20 | fail→pass | 7,145 | 5,355 | -25% | 1 | 1 | 0% | 845 | 1,197 | +42% | 0 | 0 | — |
case-21 | fail→pass | 14,868 | 9,610 | -35% | 1 | 1 | 0% | 2,193 | 1,835 | -16% | 0 | 0 | — |
case-22 | fail→pass | 6,248 | 5,395 | -14% | 1 | 1 | 0% | 733 | 1,330 | +81% | 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. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 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 +59 percentage points is the difference between those two pass rates over the 19 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.