Add a check script (scripts/check-skill-description-length.mjs) that
verifies SKILL.md description fields do not exceed 1200 characters.
Also trim redundant use-case entries in browser, chrome-bridge, and
computer-automation skill descriptions to stay within the limit.
Add testing-related trigger keywords (e.g. "test app", "QA", "verify",
"end-to-end test") to all automation skills so they are correctly matched
when users want to test applications.
Clarify browser vs desktop automation: browser skill now states it runs
headless and does not take over mouse/keyboard; computer-automation skill
warns that it does take over input devices and recommends browser skill
for web apps.
Add a new Midscene-powered skill for automating HarmonyOS NEXT devices via HDC, supporting natural language commands for taps, swipes, text input, app launches, and screenshots.
Add CRITICAL RULE #4 requiring agents to proactively report task results
before finishing. Previously, result summarization was only a soft "best
practice" recommendation, causing agents to silently end after the last
automation step — users had to send a second message to get results.
Changes across all 5 skills (browser, chrome-bridge, computer, android, ios):
- Add CRITICAL RULE #4: "Always report task results before finishing"
- Add "Report results" as a workflow step
- Strengthen Best Practice from "Summarize report files" to "Always report
results after completion" with explicit requirements
Make disconnect optional in Chrome Bridge automation to support multi-turn
conversation scenarios. Previously, the workflow mandated disconnecting
after every task, breaking the user's ability to continue interacting with
the browser in subsequent conversation turns.
Changes:
- Update Workflow Pattern to explain that browser state is always preserved
- Change disconnect from mandatory step to conditional (only when fully done)
- Update Best Practice #5 from "Disconnect when done" to "Disconnect only
when fully done" with guidance for multi-turn conversations
- Health check: observe connect command output first; only do manual
health check if connect did not already perform one
- Add best practice #8: export PATH on macOS before running to prevent
system_profiler not found errors
https://claude.ai/code/session_014boNXomePtBtKmFhA5LeFn
- Rewrite CRITICAL RULES to describe principles instead of specific parameter values
- Replace old prerequisites with proper model configuration section and examples (Gemini, Qwen3-VL via OpenRouter, Doubao Seed 1.6)
- Fix incorrect MIDSCENE_MODEL_FAMILY values (gemini, qwen3-vl, doubao-vision)
- Simplify "never run in background" best practice descriptions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
quanruzhuoxiu