github-project-discovery-hybrid

name	github-project-discovery-hybrid
description	Multi-angle GitHub project discovery using browser + GitHub API + trending, with practical clone/install/absorption workflow

GitHub Project Discovery & Absorption Hybrid Workflow

Overview

No AI halluncination about whether a project is good or not — use the browser to actually visit GitHub to check trending, topics, API search, then clone + verify via README.

Step 1: Multi-Angle Search

Run three search vectors simultaneously:

Vector A: GitHub Trending (This Week)

browser_navigate(url="https://github.com/trending?since=weekly")
browser_snapshot()

Review this week's hottest repos, filter for ones relevant to our stack.

Vector B: GitHub Topics

Browse topic pages one by one:

https://github.com/topics/<topic-name>

Common topics: wechat-miniprogram, cocos-creator, ollama, macos-automation, local-llm

Vector C: GitHub API Search (curl)

Use the API for precise searches in specific domains:

curl -s "https://api.github.com/search/repositories?q=<keyword>&sort=stars&order=desc&per_page=10"

Step 2: Filter + Verify

Don't trust descriptions blindly — read each project's README:

1. Active? — Recent commits? (within 6 months)
2. Practical? — Can it be installed and used? Any dependency issues?
3. Relevant? — Suitable for our macOS M1 / 16GB RAM / WeChat Mini Program / Cocos Creator environment?

Step 3: Take Action

For each project, decide one of the following:

• Clone + Absorb (e.g. agent-skills) — clone and read source code, patch into existing skills
• Install + Test (e.g. SkillSpector) — install and immediately test run to verify it works
• Clone but not immediately usable (e.g. ccc-devtools) — acknowledge limitations, document reasons

Step 4: Absorption (Most Important)

Don't stop at cloning — do the following:

1. Read each skill/component's README
2. Compare against existing Hermes skills to identify gaps
3. Use skill_manage(action='patch') / skill_manage(action='create') to integrate

Pitfalls

• Don't delegate_task for searching — subagents tend to be lazy and recommend mainstream projects (PyTorch, LangChain). Do your own browser research.
• Check project structure before installing tools — ccc-devtools is Cocos Creator-specific, won't work with WeChat Mini Program structure
• Token consumption management — browser navigation and curl API search are cheap, but cloning large repos and reading long READMEs burn context. Stop if no progress in 5 minutes.
• Don't re-report installed projects — use memory to track what's already installed
• 3 failures / 5 min no progress → use brake system to ask Chris
• Absorption is the heaviest work — cloning is just the start. Real time is spent on: reading each README → comparing with existing skills → deciding patch/create → writing content. Reserve sufficient token budget.
• Don't over-absorb at once — a project typically has 20+ components. Pick the 3-5 highest-value ones to absorb, record the rest for future reference.