# 策展 · X (Twitter) 🔥🔥
> 📖 本站完整內容索引（documentation index）：[llms.txt](/llms.txt)
> 作者：anita · vellum.ai 👾🦾 (@anitakirkovska) · 平台：X (Twitter) · 日期：2026-06-19
> 原始來源：https://x.com/anitakirkovska/status/2067654220027744623
## 中文摘要
# 我是如何將 MiniMax 變成 Fable 5 的（成本降低 97%）
我一直在打造自己的 MiniMax 設定，它的表現比預期中更接近 Fable 5。
其中的秘訣就在於 SKILLs 和 plugins。🥘
我也試過 Kimi K2，但無論這些基準測試（benchmarks）怎麼說，它的效能表現根本無法相提並論：
展開數據表
| Fable 5 | Opus 4.8 | Kimi K2.6 | MiniMax M3 |
|---|
| SWE-Bench Verified | ~95% | ~88% | ~80% | ~80% |
| SWE-Bench Pro | ~80% | ~70% | ~58% | ~58% |
| Terminal-Bench | ~84% | ~72% | ~67% | ~66% |
| GPQA Diamond | ~94% | ~91% | ~90% | ~93% |
| MCP-Atlas | ~83% | 無資料 | 無資料 | ~74% |
展開數據表
| Input price / 1M tokens | Output price / 1M tokens | Context window | Max output |
|---|
| Fable 5 | $10 | $50 | 1M | 128K |
| Opus 4.8 | $5 | $25 | 1M | 64K |
| Kimi K2.6 | $0.95 cache miss / $0.16 cache hit | $4.00 | 256K / 262,144 | 32K default max_tokens |
| MiniMax M3 | $0.30 <=512K promo / $0.60 >512K promo | $1.20 <=512K promo / $2.40 >512K promo | 1M | 官方文件未明確公布 |
展開畫面重點
畫面顯示一個通訊軟體中的對話串（Thread），內容如下：
- Ava (VIP APP) 於週二下午 3:41 發送：
「Meeting notes on your call: https://app.notion.com/p/Impromptu-Google-Meet-Meeting-06-16-26-3819035c57bd81899344f0e2736aea30
On the tasks for me:
• Create Notion summary re: plugins + best practices; Slack link to Anita
Where should I send this? (Slack / Email / Notion)」
- anita (📅) 於週二下午 3:42 回覆：
「hey @Ava send this in this thread as a Notion link once you're done with it」
- Ava (VIP APP) 於週二下午 3:45 回覆：
「Done — Notion (you'll find it under Ava > Insights and ideas): Plugins & plugin best practices (Jun 16, 2026)」
畫面重點在於展示 AI 助理如何協助整理會議筆記、執行指定任務，並在協作平台中與使用者進行互動與確認進度。
展開畫面重點
Meta-CLI
🛠️ A skill for shipping CLI tools + their agent-facing SKILL.md the right way.
Most public CLI repos hand you a binary and a half-written README. Meta-CLI is the missing layer above: a meta-skill that another AI agent (Claude, GPT, Vellum, anything) reads before it builds the next kalshi-cli or elevenlabs-cli.
If you're an agent picking up "build me a CLI for service X," start with SKILL.md.
🧭 What's in here
📦 One skill, the entire pattern. The decision tree, the CLI anatomy (subcommands, error model, --local mode), the SKILL.md format, the two-repo hygiene rules, the pre-flight checklist that catches the bugs v1 ships with, and the common-mistakes list distilling what other public CLIs got wrong.
🎯 Built for agents picking up concrete tasks. The skill isn't theory. It's a checklist another agent follows when an external service needs a generic, installable wrapper.
🚀 Generic install, generic look, generic feel. pip install git+https://github.com/AnitaKirkovska/Meta-CLI.git (or any clone path). Zero personal content in this repo. Zero proprietary bits. Strict separation from any private usage.
展開畫面重點
畫面左側為通訊軟體的導覽列，包含以下項目：
- Ava
- Team
- Marketing
- Direct messages
- Priority
- Social
- Customers
畫面中央彈出一個標題為「Read conversations」的視窗，列出了與「Ava APP」相關的頻道列表：
- ava-agent-cup（鎖頭圖示）
- ava-meeting-notes（鎖頭圖示）
- ava-misc（井字號圖示）
- ava-product（井字號圖示）
- ava-signups（鎖頭圖示）
- ava-socials（鎖頭圖示）
右上角可見部分文字內容：「Austria v Jordan: Ava nicks...」以及「more likely land at 10 decads」。該介面呈現了典型的團隊協作通訊工具佈局，透過鎖頭圖示區分私人或受限頻道，井字號則代表公開頻道。