convez.ai
Skill
GiaSip

giasip-skills

GiaSip/giasip-skills

Ver no GitHub
InicialPluginScore · de 0 a 100

research & dispatch skills for Claude Code — 研究调度(Quick Recon + fact-check)+ 多模型调用器(纯调用器版)

Instalação

$npx convez-ai GiaSip/giasip-skills

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
13
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

giasip-skills

Version License Claude Code

Custom Claude Code skills by GiaSip · github.com/GiaSip

Claude Code-native. A vendor-neutral portable-playbook version (usable from Codex and other agents) is maintained separately.

SkillDescription
giasip-researchResearch orchestrator — runs a Quick Recon with SubAgents first to map the landscape and knowledge gaps, then decides whether to escalate to an external Deep Research platform. Built-in two-round Recon, Claim Ledger quality gate, and independent fact-check protocol ensure external platforms only handle what truly needs deep digging.
giasip-dispatchMulti-model dispatcher — sends a task or prompt to other AI models (Codex / Gemini / Kimi / DeepSeek / Doubao / Qwen / GLM / MiniMax) and retrieves results. Includes complexity routing guidelines to help pick the right dispatch strategy (API vs CLI vs SubAgent, single vs multi), but the final model choice is left to your Claude's judgment.

Skill Structure

skills/
├── giasip-research/
│   ├── SKILL.md                          # Core skill definition
│   └── references/
│       ├── platform-profiles.md          # Deep Research platform capability cards
│       ├── matching-rules.md             # Platform matching decision tree
│       ├── fact-check-protocol.md        # Fact-check protocol (v2.2+v2.4) + Mini Assurance
│       └── subagent-templates.md         # SubAgent instruction templates + unit sanity check
│
└── giasip-dispatch/
    ├── SKILL.md                          # Core skill definition
    ├── scripts/
    │   ├── api-dispatch.sh               # API direct call (DeepSeek/Qwen/GLM/Doubao/MiniMax)
    │   ├── codex-appserver.mjs           # Codex App Server protocol (no cold start)
    │   ├── gemini-supervisor.sh          # Gemini CLI + retry/fallback/circuit breaker
    │   ├── kimi-dispatch.sh              # Kimi dispatch + thinking mode control
    │   ├── dispatch-persist.mjs          # Response logging sink
    │   └── stop-review-gate.mjs          # Codex stop-hook for code review gating
    └── references/
        └── model-roster.md              # Full model roster + multi-dispatch lineup guide

Installation

Option 1: npx skills add (recommended, one-liner)

# Install all skills globally (to ~/.claude/skills/)
npx skills add GiaSip/giasip-skills -g --all

# Install a single skill
npx skills add GiaSip/giasip-skills -g --skill giasip-research

# List available skills in this repo
npx skills add GiaSip/giasip-skills -l

Option 2: Claude Code plugin

/plugin marketplace add GiaSip/giasip-skills
/plugin install giasip-skills@giasip-skills

Option 3: git clone

git clone https://github.com/GiaSip/giasip-skills
cp -R giasip-skills/skills/giasip-research ~/.claude/skills/giasip-research
cp -R giasip-skills/skills/giasip-dispatch ~/.claude/skills/giasip-dispatch

All three methods register the same slash commands: /giasip-research and /giasip-dispatch. You can also trigger them by describing your intent (e.g., "research X for me" or "run this with Kimi"). The dispatch scripts live in the skill's own scripts/ subdirectory; the dispatch SKILL.md has the agent set BASE_DIR to that directory once at the start of a session, so the script calls work whether installed under ~/.claude/skills/ or in a plugin cache directory.


Quick Start — Dispatch in 5 Minutes

  1. Set up one API key (DeepSeek is a good start):

    mkdir -p ~/.config/ai-keys
    echo 'export DEEPSEEK_API_KEY=sk-...' > ~/.config/ai-keys/deepseek.env
    
  2. Install the skills:

    npx skills add GiaSip/giasip-skills -g --all
    
  3. Try it — in Claude Code, just say:

    • "Run this analysis with DeepSeek: what are the tradeoffs between PostgreSQL and MongoDB?"
    • "Research the current state of humanoid robot regulations"

giasip-research — Dependencies

Near-zero external dependencies — works out of the box. Primarily uses Claude Code's built-in WebSearch / WebFetch / SubAgent. (Firecrawl can optionally serve as a fallback when WebFetch hits JS-rendered pages, but is not required.)

The only setup needed: fill in the platform availability table in skills/giasip-research/references/platform-profiles.md with your actual Deep Research subscriptions (ChatGPT / Gemini / Perplexity / Kimi, etc.) — the matching logic uses this to skip unsubscribed platforms.

giasip-dispatch — Dependencies

Two types of dispatch channels; configure what you need:

1. API direct call (just needs an API key — fastest)

Supports DeepSeek / Qwen / GLM / Doubao / MiniMax. Place the corresponding .env file in ~/.config/ai-keys/:

ModelFileContent
DeepSeekdeepseek.envexport DEEPSEEK_API_KEY=...
Qwen (Tongyi)dashscope.envexport DASHSCOPE_API_KEY=...
GLM (Zhipu)zai.envexport ZAI_API_KEY=...
Doubao (Volcengine)volcengine.envexport ARK_API_KEY=...
MiniMaxminimax.envexport MINIMAX_API_KEY=...

Test (adjust the path to your install location — e.g. a global install): ~/.claude/skills/giasip-dispatch/scripts/api-dispatch.sh --model deepseek "Hello"

Specific model names (e.g., deepseek-v4-pro) are defined in the case branches of api-dispatch.sh and may change as vendors release new versions — update MODEL_ID in the script if a call fails. See references/model-roster.md for the current roster.

2. CLI invocation (requires local install + login)

ModelInstallAuth
Codexnpm i -g @openai/codexChatGPT account
Gemininpm i -g @google/gemini-cliGoogle account
Kimiuv tool install kimi-cli (or API key only)kimi.com / Moonshot key

Dependency check: command -v codex gemini kimi node curl python3 jq perl

Kimi has two backends. The default kimi-dispatch.sh calls the Moonshot API directly (it's really an API channel, not CLI) — needs ~/.config/ai-keys/kimi-moonshot.env containing MOONSHOT_API_KEY, no kimi CLI required. Adding KIMI_FOR_CODING=1 switches to the Kimi CLI coding endpoint — that path needs the kimi CLI installed plus ~/.config/ai-keys/kimi.env containing KIMI_API_KEY.

perl is required by the Gemini and Kimi wrappers for portable timeout control — preinstalled on macOS, but install it on minimal Linux images.

All scripts read keys via source ~/.config/ai-keys/*.env — your keys stay local and are never in this repo.


What's Included

FileDescription
skills/giasip-research/references/platform-profiles.mdDeep Research platform capability cards (speed/quality/context ratings)
skills/giasip-research/references/matching-rules.mdPlatform matching decision tree (language routing, special requirements)
skills/giasip-research/references/fact-check-protocol.mdIndependent fact-check protocol with cross-faction discipline
skills/giasip-research/references/subagent-templates.mdSubAgent instruction templates with ClaimCard schema
skills/giasip-dispatch/references/model-roster.mdModel roster with per-model strengths and multi-dispatch lineups
skills/giasip-dispatch/scripts/dispatch-persist.mjsPersists dispatch responses to ~/.cache/dispatch/ (call explicitly or hook into dispatch scripts)
skills/giasip-dispatch/scripts/stop-review-gate.mjsClaude Code stop hook — advisory Codex code review gate

Chinese version

A Chinese version of all documentation is available under locales/zh/.

Changelog

See CHANGELOG.md for version history.

License

MIT © GiaSip

Ver o original no GitHub

Bom para

Relacionados