0:00.000–0:01.676
So Kimi K3 just got released and
0:01.676–0:04.320
it's the number one front-end model on the arena.
0:04.620–0:07.380
I'm going to show you exactly how you can set it up in your clawed code
0:07.380–0:10.220
and how you can make this viral visual in under 10 minutes.
0:11.720–0:13.060
So this is Kimi K3.
0:13.520–0:16.920
It's the latest open source model from Moonshot AI Labs.
0:17.420–0:19.420
And this video on the right here is their launch video,
0:19.560–0:23.540
which they said was entirely edited and created by Kimi K3 itself.
0:23.940–0:27.360
In this video, we're going to cover what exactly Kimi K3 is.
0:28.260–0:30.220
We're going to set it up in clawed code
0:30.220–0:35.180
and then we're going to recreate this viral style templated video on the right here
0:35.180–0:40.280
where you can compare Kimi K3, Fable 5 and GPT 5.6 Sol,
0:40.760–0:42.840
giving different prompts to C-Dance.
0:44.860–0:49.700
So it's a 2.8 trillion parameter mixture of experts model
0:49.700–0:51.840
with a 1 million context window.
0:52.440–0:54.920
And the pricing is equivalent to Sonnet, which is pretty insane
0:54.920–0:58.760
because Kimi K2, K3 consistently outperforms Opus 4.
0:58.760–1:00.040
8 in a lot of tasks.
1:00.760–1:02.180
This is front-end code arena.
1:02.660–1:06.100
It ranks models based on blind votes on real front-end tasks.
1:06.540–1:08.640
So people compare two outputs and then pick one.
1:08.800–1:10.940
So it's not a lab that's grading its own homework.
1:11.600–1:15.720
And you can see row 18, we had the old Kimi K2.6 model.
1:16.280–1:18.400
And now it's all the way up in first place.
1:18.940–1:21.580
So they went from 18th to first place in one release.
1:21.580–1:25.300
So if you use models for design work, for UI,
1:25.580–1:27.740
pretty much anything that a person ends up looking at,
1:27.980–1:29.780
this is the chart that matters most to you.
1:30.160–1:31.700
That's just the front-end board though.
1:31.880–1:33.684
And I don't want to oversell it because
1:33.684–1:35.940
Kimi K3 actually does sit around the third place
1:35.940–1:39.417
and the second place on most of the general coding and
1:39.417–1:40.840
agentic benchmarks.
1:40.840–1:44.580
So here you can see that above it most of the time is Fable 5.
1:45.260–1:48.540
And it does beat GPT 5.6 a lot of the time as well.
1:49.060–1:52.400
And honestly, this is the part that puts real pressure on Anthropic and OpenAI
1:52.400–1:57.080
because being this close to the top used to mean that you had to pay huge premium for it.
1:57.500–2:00.860
And now you can get a model that's the equivalent to GPT 5.6
2:00.860–2:04.120
and even better than Opus 4.8 for the price of Sonnet.
2:04.120–2:08.180
Before I show you how you can set this model up,
2:08.260–2:10.300
I'll just go through a few of the catches.
2:11.120–2:12.780
Firstly, there's no open waits yet.
2:13.100–2:16.460
They've said that they're going to release the open waits on July 27th.
2:16.460–2:19.440
But as of today, there's nothing public yet.
2:20.240–2:22.660
Another one is that it's slow on hard tasks.
2:22.840–2:26.720
So a user reported that it took up to 35 minutes for a single front-end task.
2:27.320–2:29.620
And then the third one is that it trains on your traffic.
2:29.700–2:31.500
So anything that you're sending it through the API,
2:31.500–2:34.660
it's going to train its future models on that.
2:35.080–2:36.340
And it is a Chinese model.
2:36.540–2:40.200
So you never know what's going to happen with that data.
2:41.360–2:41.800
All right.
2:41.800–2:43.120
So let's jump into the setup.
2:43.860–2:44.020
All right.
2:44.020–2:45.500
So the first step is getting our key.
2:45.940–2:48.040
And to get our key, we're going to go to Kimi.com.
2:48.620–2:49.700
You're going to want to log in.
2:49.800–2:51.940
And then you're going to need to get on the cheapest plan.
2:52.020–2:53.760
So just come down to your name down the bottom here.
2:54.100–2:55.120
Click on upgrade plan.
2:55.580–2:57.020
And then I've just chosen this plan here,
2:57.060–2:58.200
but you can choose any of your plans.
2:58.380–2:59.340
They all come with K3.
2:59.340–3:01.720
And this beats out using the API,
3:02.040–3:03.400
because if you're going to use the API,
3:03.480–3:05.420
you're going to have to pay for that usage fee,
3:05.500–3:08.440
which is going to end up being more than if you're just going to use the plan.
3:10.020–3:12.020
Now that we're logged in and subscribed to our plan,
3:12.080–3:15.100
we need to go to the console page to get our API key.
3:15.180–3:16.340
So I'm going to go to Kimi code.
3:17.140–3:18.840
Then I'm going to go to console.
3:19.320–3:21.040
Next, we just want to create an API key.
3:21.420–3:22.800
So just name it whatever you want.
3:22.980–3:23.620
Click create.
3:24.020–3:25.900
Then you need to copy this key here.
3:25.900–3:28.500
And we're just going to take this key into Claude code now.
3:29.000–3:29.180
All right.
3:29.180–3:30.040
So I've got my key copied.
3:30.180–3:32.960
And now I'm in VS code with the Claude code plugin installed.
3:33.220–3:35.540
I'm just going to paste the key in here and say,
3:35.720–3:39.960
please set up Kimi K3 for me.
3:40.040–3:43.480
So I can use it as a sub agent.
3:43.820–3:44.940
One thing to know though,
3:44.940–3:47.060
is that an API key is essentially a password.
3:47.760–3:50.680
And I've just pasted this one in here because it's just a $20 burner.
3:50.680–3:53.540
And I can cancel that key anytime on their site.
3:53.820–3:59.240
So I just wouldn't paste any important keys that matter directly into your Claude code chat.
3:59.520–4:03.000
I definitely put them straight into your .emv file.
4:03.600–4:03.780
All right.
4:03.800–4:06.540
So Claude just got back to me about two minutes later saying,
4:06.660–4:08.540
Kimi K3 sub agent is done.
4:08.820–4:10.080
TLDR, Kimi builder.
4:10.280–4:13.220
A sub agent is set up and the key is verified working.
4:13.380–4:15.360
New agent types load at session start.
4:15.460–4:17.520
So it becomes spawnable after your next slash clear.
4:17.520–4:19.280
Now I'm just going to paste this in and say,
4:20.020–4:24.160
I'm going to say, how can we use this sub agent?
4:25.440–4:28.320
Kimi is now set up fully in my Claude code as a sub agent.
4:28.460–4:29.580
So whenever I want to use Kimi,
4:29.880–4:32.880
I'll just tell Claude to use that Kimi sub agent.
4:33.300–4:36.140
So now we'll jump into creating an actual test.
4:36.440–4:37.720
So we're going to use this example here,
4:37.780–4:40.160
which is through the Wireflow API.
4:40.620–4:43.360
And Wireflow is just a creative node-based canvas
4:43.360–4:46.760
that lets us add prompts and then connect those prompts
4:46.760–4:48.340
to video models or image models.
4:48.580–4:52.780
So for this example, we're going to have Kimi create a creative prompt
4:52.780–4:54.680
based on this still image here.
4:55.360–4:57.800
And then we're going to have Fable 5 create a prompt
4:57.800–5:00.100
as well as GPT 5.6 create a prompt.
5:00.300–5:03.580
And then we're going to see what those different outputs
5:03.580–5:04.920
look like side by side.
5:07.080–5:07.540
All right.
5:07.540–5:09.840
Next steps are we just have to set up the Wireflow API.
5:09.840–5:13.020
So I'm just going to come to wireflow.ai slash API
5:13.020–5:14.740
and I'm going to click get the key.
5:15.040–5:17.120
And then it's going to take me to slash settings.
5:17.440–5:19.960
And then in here, I'm going to click on API keys
5:19.960–5:21.880
and then click on create a new key.
5:22.060–5:25.180
Just name it Kimi create key, copy this key.
5:25.280–5:26.980
And we're just going to do the exact same process
5:26.980–5:28.800
as before with the Kimi key.
5:28.900–5:30.360
So I'm just going to paste it in here and say,
5:30.500–5:35.780
please set up my API key for Wireflow, Wireflow's API.
5:36.400–5:36.820
All right.
5:36.860–5:38.440
So our Wireflow key is set up now
5:38.440–5:40.140
and it's confirmed working.
5:40.300–5:41.680
So now we're going to go back to Wireflow.
5:41.940–5:43.460
This is the template that we're using.
5:43.700–5:46.360
So it's a one start frame, three AI directors,
5:46.680–5:49.520
Kimi K3, Claude Fable 5 and GPT 5.6.
5:49.820–5:51.280
So once you have everything set up,
5:51.320–5:52.420
you just click on use this template
5:52.420–5:53.720
and then it will open the template,
5:53.860–5:54.720
which looks like this.
5:54.860–5:56.760
And you can actually run the workflow from in here,
5:56.840–6:00.140
but we're going to use Claude code to write all the prompts,
6:00.300–6:01.560
inject the prompts in the workflow,
6:01.560–6:02.820
and then we'll come back to the workflow
6:02.820–6:05.120
and run the actual C dance nodes one at a time.
6:05.500–6:07.080
Because we have Claude code already connected
6:07.080–6:08.180
to our Wireflow account.
6:08.300–6:09.880
We can just grab this URL here,
6:10.160–6:12.080
come back to our Claude code with Kimi
6:12.080–6:13.600
and just paste the URL and say,
6:13.760–6:16.760
can you learn this workflow?
6:17.400–6:21.820
I'm going to want to change the prompts.
6:22.360–6:22.480
Okay.
6:22.560–6:23.200
So it said,
6:23.740–6:24.320
TLDR,
6:24.320–6:26.700
it's a three-way AI directors comparison flow,
6:27.160–6:29.800
one shared start frame fans out into three C dance,
6:29.900–6:30.380
two nodes,
6:30.500–6:32.340
each driven by a different model's motion prompt.
6:32.780–6:32.960
All right.
6:32.960–6:33.960
So next we're going to write,
6:34.380–6:38.420
please spawn three different sub agents,
6:38.940–6:40.700
one Kimi K3,
6:41.220–6:42.380
one Fable,
6:42.380–6:45.400
and one GPT 5.6,
6:45.540–6:49.840
and have them each look at the reference image,
6:49.980–6:53.640
then create their own motion prompt
6:53.640–6:58.760
and insert it into the relevant text box,
6:58.980–6:59.360
please.
6:59.780–7:01.880
So Claude code has just replied and it said,
7:01.980–7:02.140
done.
7:02.220–7:04.120
All three models view the actual start frame,
7:04.480–7:05.640
wrote their own motion prompt,
7:05.640–7:07.440
and the prompts are live within the flow.
7:07.440–7:09.260
So you can see Kimi K3,
7:09.320–7:10.180
110 words,
7:10.280–7:13.360
and then it kind of describes the different actions
7:13.360–7:15.160
and motions that each one of them took here.
7:15.320–7:16.760
And so now we can go back into the flow
7:16.760–7:19.120
and you can see that this has already changed.
7:19.340–7:21.600
Each prompt has been automatically inserted
7:21.600–7:22.900
into their prompt box here.
7:23.100–7:25.520
And then you can see these purple lines here.
7:25.640–7:26.600
This is the text connector.
7:26.740–7:29.420
So it's connecting into the C dance to here.
7:29.820–7:32.200
And then this start frame is also getting connected.
7:32.200–7:35.460
So each of these videos starts from the same start frame.
7:35.460–7:39.380
Then the language models can take it in different directions.
7:39.860–7:41.020
And so I've already run these models
7:41.020–7:42.840
and this was the output.
7:45.840–7:47.380
And the next steps in this workflow
7:47.380–7:49.060
is essentially a image compositor.
7:49.160–7:51.840
So it's just going to use these black overlays
7:51.840–7:54.740
and then it's sending this into a video editor.
7:55.000–7:57.180
And this is essentially just putting those overlays
7:57.180–7:59.660
on those C dance to clips here
7:59.660–8:02.380
so that we can post this on Twitter or LinkedIn.
8:02.380–8:05.840
So everything I've shown in this video so far
8:05.840–8:06.980
I'll include it in the description
8:06.980–8:08.380
including this template here
8:08.380–8:11.300
which also includes a skill in it
8:11.300–8:14.040
to set up those sub-agents in Claude Code.
8:14.900–8:17.360
So hopefully you guys found this tutorial useful
8:17.360–8:19.960
and you're able to set up Kimi K3 for yourselves.
8:19.960–8:20.000
Let's see.
0:00.000–0:01.676
So Kimi K3 just got released and
所以Kimi K3剛剛發布,並且
0:01.676–0:04.320
it's the number one front-end model on the arena.
它已經是Arena上排名第一的前端模型。
0:04.620–0:07.380
I'm going to show you exactly how you can set it up in your clawed code
我將向你展示如何在Clawed Code中設置它,
0:07.380–0:10.220
and how you can make this viral visual in under 10 minutes.
以及如何在10分鐘內製作這個病毒式視覺效果。
0:11.720–0:13.060
So this is Kimi K3.
所以這是Kimi K3。
0:13.520–0:16.920
It's the latest open source model from Moonshot AI Labs.
它是Moonshot AI Labs最新的開源模型。
0:17.420–0:19.420
And this video on the right here is their launch video,
而右邊的這段視頻是他們的發布視頻,
0:19.560–0:23.540
which they said was entirely edited and created by Kimi K3 itself.
據他們說,該視頻完全由Kimi K3本身編輯和創建。
0:23.940–0:27.360
In this video, we're going to cover what exactly Kimi K3 is.
在本視頻中,我們將介紹Kimi K3究竟是什麼。
0:28.260–0:30.220
We're going to set it up in clawed code
我們將把它設置在Clawed Code中,
0:30.220–0:35.180
and then we're going to recreate this viral style templated video on the right here
然後我們將重製右邊這段病毒式風格模板視頻,
0:35.180–0:40.280
where you can compare Kimi K3, Fable 5 and GPT 5.6 Sol,
你可以在其中比較Kimi K3、Fable 5和GPT 5.6 Sol,
0:40.760–0:42.840
giving different prompts to C-Dance.
並向C-Dance提供不同的提示。
0:44.860–0:49.700
So it's a 2.8 trillion parameter mixture of experts model
所以它是一個擁有2.8萬億參數的專家混合模型,
0:49.700–0:51.840
with a 1 million context window.
並具備100萬的上下文窗口。
0:52.440–0:54.920
And the pricing is equivalent to Sonnet, which is pretty insane
並且其定價與Sonnet相當,這相當驚人,
0:54.920–0:58.760
because Kimi K2, K3 consistently outperforms Opus 4.
因為Kimi K2、K3在許多任務中持續優於Opus 4.8。
0:58.760–1:00.040
8 in a lot of tasks.
這是前端代碼Arena。
1:00.760–1:02.180
This is front-end code arena.
它根據真實前端任務的盲選投票對模型進行排名。
1:02.660–1:06.100
It ranks models based on blind votes on real front-end tasks.
所以人們比較兩個輸出結果,然後選擇其中一個。
1:06.540–1:08.640
So people compare two outputs and then pick one.
所以這不是實驗室在批改自己的作業。
1:08.800–1:10.940
So it's not a lab that's grading its own homework.
你可以看到第18行,我們有舊的Kimi K2.6模型。
1:11.600–1:15.720
And you can see row 18, we had the old Kimi K2.6 model.
而現在它已經上升到第一名。
1:16.280–1:18.400
And now it's all the way up in first place.
所以他們在一次發布中從第18名躍升至第一名。
1:18.940–1:21.580
So they went from 18th to first place in one release.
所以如果你使用模型進行設計工作、UI設計,
1:21.580–1:25.300
So if you use models for design work, for UI,
基本上是人最終會看到的一切,
1:25.580–1:27.740
pretty much anything that a person ends up looking at,
這張圖表對你最為重要。
1:27.980–1:29.780
this is the chart that matters most to you.
這只是前端榜單。
1:30.160–1:31.700
That's just the front-end board though.
我不想過度推銷它,因為
1:31.880–1:33.684
And I don't want to oversell it because
Kimi K3實際上在大多數通用編程和
1:33.684–1:35.940
Kimi K3 actually does sit around the third place
智能體基準測試中排名第三和第二。
1:35.940–1:39.417
and the second place on most of the general coding and
所以你可以在這裡看到,大部分時間排在它上面的是Fable 5。
1:39.417–1:40.840
agentic benchmarks.
而且它確實也經常擊敗GPT 5.6。
1:40.840–1:44.580
So here you can see that above it most of the time is Fable 5.
老實說,這部分是真正給Anthropic和OpenAI帶來壓力的原因,
1:45.260–1:48.540
And it does beat GPT 5.6 a lot of the time as well.
因為曾經接近頂端意味著你必須為此支付巨大的溢價。
1:49.060–1:52.400
And honestly, this is the part that puts real pressure on Anthropic and OpenAI
而現在你可以獲得一個與GPT 5.6相當
1:52.400–1:57.080
because being this close to the top used to mean that you had to pay huge premium for it.
甚至優於Opus 4.8的模型,價格卻僅為Sonnet的水平。
1:57.500–2:00.860
And now you can get a model that's the equivalent to GPT 5.6
在我向你展示如何設置這個模型之前,
2:00.860–2:04.120
and even better than Opus 4.8 for the price of Sonnet.
而且價格僅相當於 Sonnet,性能甚至優於 Opus 4.8。
2:04.120–2:08.180
Before I show you how you can set this model up,
在展示如何設定這個模型之前,
2:08.260–2:10.300
I'll just go through a few of the catches.
接下來我會說明幾個需要注意的地方。
2:11.120–2:12.780
Firstly, there's no open waits yet.
首先,目前還沒有開放等待名單。
2:13.100–2:16.460
They've said that they're going to release the open waits on July 27th.
他們表示將在 7 月 27 日開放等待名單。
2:16.460–2:19.440
But as of today, there's nothing public yet.
但截至今天,還沒有任何公開資訊。
2:20.240–2:22.660
Another one is that it's slow on hard tasks.
另一個問題是,在處理複雜任務時速度較慢。
2:22.840–2:26.720
So a user reported that it took up to 35 minutes for a single front-end task.
有使用者回報,單一前端任務最多花了 35 分鐘。
2:27.320–2:29.620
And then the third one is that it trains on your traffic.
第三點是,它會利用你的流量進行訓練。
2:29.700–2:31.500
So anything that you're sending it through the API,
也就是說,你透過 API 傳送給它的所有內容,
2:31.500–2:34.660
it's going to train its future models on that.
都會被用來訓練它未來的模型。
2:35.080–2:36.340
And it is a Chinese model.
而且這是一個中國模型。
2:36.540–2:40.200
So you never know what's going to happen with that data.
所以你永遠不知道這些資料會發生什麼事。
2:41.360–2:41.800
All right.
好了。
2:41.800–2:43.120
So let's jump into the setup.
現在讓我們進入設定步驟。
2:43.860–2:44.020
All right.
好了。
2:44.020–2:45.500
So the first step is getting our key.
第一步是取得我們的金鑰。
2:45.940–2:48.040
And to get our key, we're going to go to Kimi.com.
要取得金鑰,我們會前往 Kimi.com。
2:48.620–2:49.700
You're going to want to log in.
你需要登入。
2:49.800–2:51.940
And then you're going to need to get on the cheapest plan.
然後你需要選擇最便宜的方案。
2:52.020–2:53.760
So just come down to your name down the bottom here.
所以請往下滑到這裡底部的你的名稱。
2:54.100–2:55.120
Click on upgrade plan.
點擊「升級方案」。
2:55.580–2:57.020
And then I've just chosen this plan here,
然後我這裡選擇了這個方案,
2:57.060–2:58.200
but you can choose any of your plans.
但你可以選擇任何方案。
2:58.380–2:59.340
They all come with K3.
所有方案都包含 K3。
2:59.340–3:01.720
And this beats out using the API,
這比使用 API 更好,
3:02.040–3:03.400
because if you're going to use the API,
因為如果你要使用 API,
3:03.480–3:05.420
you're going to have to pay for that usage fee,
你就必須支付使用費,
3:05.500–3:08.440
which is going to end up being more than if you're just going to use the plan.
這最終會比直接使用方案更昂貴。
3:10.020–3:12.020
Now that we're logged in and subscribed to our plan,
現在我們已登入並訂閱了方案,
3:12.080–3:15.100
we need to go to the console page to get our API key.
我們需要前往主控台頁面以取得 API 金鑰。
3:15.180–3:16.340
So I'm going to go to Kimi code.
所以我會前往 Kimi code。
3:17.140–3:18.840
Then I'm going to go to console.
然後我會前往主控台。
3:19.320–3:21.040
Next, we just want to create an API key.
接下來,我們只需要建立一個 API 金鑰。
3:21.420–3:22.800
So just name it whatever you want.
所以請隨意命名。
3:22.980–3:23.620
Click create.
點擊建立。
3:24.020–3:25.900
Then you need to copy this key here.
然後你需要複製這裡的金鑰。
3:25.900–3:28.500
And we're just going to take this key into Claude code now.
我們將把這個金鑰帶入 Claude code。
3:29.000–3:29.180
All right.
好了。
3:29.180–3:30.040
So I've got my key copied.
所以我已經複製了我的金鑰。
3:30.180–3:32.960
And now I'm in VS code with the Claude code plugin installed.
現在我在 VS code 中,並已安裝 Claude code 外掛。
3:33.220–3:35.540
I'm just going to paste the key in here and say,
我只會把金鑰貼在這裡並說,
3:35.720–3:39.960
please set up Kimi K3 for me.
請幫我設定 Kimi K3。
3:40.040–3:43.480
So I can use it as a sub agent.
這樣我就能把它當作子代理(sub agent)來使用。
3:43.820–3:44.940
One thing to know though,
不過有一件事需要了解,
3:44.940–3:47.060
is that an API key is essentially a password.
那就是 API 金鑰本質上就相當於密碼。
3:47.760–3:50.680
And I've just pasted this one in here because it's just a $20 burner.
我剛才把它貼在這裡,是因為這只是一個 20 美元的消耗型金鑰。
3:50.680–3:53.540
And I can cancel that key anytime on their site.
而且我可以隨時在他們的網站上取消該金鑰。
3:53.820–3:59.240
So I just wouldn't paste any important keys that matter directly into your Claude code chat.
所以我不會把任何重要的金鑰直接貼到 Claude Code 的聊天視窗中。
3:59.520–4:03.000
I definitely put them straight into your .emv file.
我絕對會把金鑰直接寫進你的 .env 檔案裡。
4:03.600–4:03.780
All right.
好的。
4:03.800–4:06.540
So Claude just got back to me about two minutes later saying,
大約兩分鐘後,Claude 回覆我說,
4:06.660–4:08.540
Kimi K3 sub agent is done.
Kimi K3 子代理已設定完成。
4:08.820–4:10.080
TLDR, Kimi builder.
簡短來說,Kimi 建構器。
4:10.280–4:13.220
A sub agent is set up and the key is verified working.
子代理已設定,且金鑰驗證運作正常。
4:13.380–4:15.360
New agent types load at session start.
新的代理類型會在工作階段開始時載入。
4:15.460–4:17.520
So it becomes spawnable after your next slash clear.
所以在你下次執行 /clear 指令後,它就可以被啟動。
4:17.520–4:19.280
Now I'm just going to paste this in and say,
現在我要把這個貼上去,並說,
4:20.020–4:24.160
I'm going to say, how can we use this sub agent?
我要問說,我們該如何使用這個子代理?
4:25.440–4:28.320
Kimi is now set up fully in my Claude code as a sub agent.
Kimi 現在已在我的 Claude Code 中完全設定為子代理。
4:28.460–4:29.580
So whenever I want to use Kimi,
所以每當我想使用 Kimi 時,
4:29.880–4:32.880
I'll just tell Claude to use that Kimi sub agent.
我就會指示 Claude 使用那個 Kimi 子代理。
4:33.300–4:36.140
So now we'll jump into creating an actual test.
現在我們來建立一個實際的測試。
4:36.440–4:37.720
So we're going to use this example here,
我們將使用這裡的範例,
4:37.780–4:40.160
which is through the Wireflow API.
它是透過 Wireflow API 進行。
4:40.620–4:43.360
And Wireflow is just a creative node-based canvas
而 Wireflow 是一個創意的節點式畫布,
4:43.360–4:46.760
that lets us add prompts and then connect those prompts
讓我們可以新增提示,然後將這些提示
4:46.760–4:48.340
to video models or image models.
連結到影片模型或圖像模型。
4:48.580–4:52.780
So for this example, we're going to have Kimi create a creative prompt
因此,在這個範例中,我們將讓 Kimi 根據這裡的靜態影像建立一個創意提示。
4:52.780–4:54.680
based on this still image here.
基於這張靜態影像。
4:55.360–4:57.800
And then we're going to have Fable 5 create a prompt
接著我們將讓 Fable 5 建立提示,
4:57.800–5:00.100
as well as GPT 5.6 create a prompt.
以及 GPT 5.6 建立提示。
5:00.300–5:03.580
And then we're going to see what those different outputs
然後我們將並排查看這些不同輸出的
5:03.580–5:04.920
look like side by side.
外觀。
5:07.080–5:07.540
All right.
好的。
5:07.540–5:09.840
Next steps are we just have to set up the Wireflow API.
下一步,我們只需要設定 Wireflow API。
5:09.840–5:13.020
So I'm just going to come to wireflow.ai slash API
所以我會前往 wireflow.ai 的 API 頁面,
5:13.020–5:14.740
and I'm going to click get the key.
並點擊取得金鑰。
5:15.040–5:17.120
And then it's going to take me to slash settings.
接著它會帶我進入設定頁面。
5:17.440–5:19.960
And then in here, I'm going to click on API keys
然後在這裡,我會點擊 API 金鑰,
5:19.960–5:21.880
and then click on create a new key.
接著點擊建立新金鑰。
5:22.060–5:25.180
Just name it Kimi create key, copy this key.
將其命名為 Kimi create key,並複製這個金鑰。
5:25.280–5:26.980
And we're just going to do the exact same process
接下來,我們將執行與之前相同的流程
5:26.980–5:28.800
as before with the Kimi key.
使用 Kimi 的 API 金鑰。
5:28.900–5:30.360
So I'm just going to paste it in here and say,
所以我只會把它貼在這裡,然後說,
5:30.500–5:35.780
please set up my API key for Wireflow, Wireflow's API.
請為 Wireflow 的 API 設定我的 API 金鑰。
5:36.400–5:36.820
All right.
好的。
5:36.860–5:38.440
So our Wireflow key is set up now
現在我們的 Wireflow 金鑰已設定完成
5:38.440–5:40.140
and it's confirmed working.
並確認運作正常。
5:40.300–5:41.680
So now we're going to go back to Wireflow.
所以現在我們要回到 Wireflow。
5:41.940–5:43.460
This is the template that we're using.
這是我們使用的範本。
5:43.700–5:46.360
So it's a one start frame, three AI directors,
它包含一個起始幀、三個 AI 導演,
5:46.680–5:49.520
Kimi K3, Claude Fable 5 and GPT 5.6.
分別是 Kimi K3、Claude Fable 5 和 GPT 5.6。
5:49.820–5:51.280
So once you have everything set up,
所以一旦你完成所有設定,
5:51.320–5:52.420
you just click on use this template
只需點擊「使用此範本」
5:52.420–5:53.720
and then it will open the template,
然後它就會開啟範本,
5:53.860–5:54.720
which looks like this.
看起來就像這樣。
5:54.860–5:56.760
And you can actually run the workflow from in here,
你實際上可以從這裡執行工作流程,
5:56.840–6:00.140
but we're going to use Claude code to write all the prompts,
但我們將使用 Claude Code 來撰寫所有提示詞,
6:00.300–6:01.560
inject the prompts in the workflow,
將提示詞注入工作流程,
6:01.560–6:02.820
and then we'll come back to the workflow
然後我們會回到工作流程
6:02.820–6:05.120
and run the actual C dance nodes one at a time.
並一次執行一個 C 舞蹈節點。
6:05.500–6:07.080
Because we have Claude code already connected
因為我們已經將 Claude Code 連接
6:07.080–6:08.180
to our Wireflow account.
到我們的 Wireflow 帳戶。
6:08.300–6:09.880
We can just grab this URL here,
我們可以複製這裡的 URL,
6:10.160–6:12.080
come back to our Claude code with Kimi
回到我們的 Claude Code 與 Kimi,
6:12.080–6:13.600
and just paste the URL and say,
然後貼上 URL 並說,
6:13.760–6:16.760
can you learn this workflow?
你能學習這個工作流程嗎?
6:17.400–6:21.820
I'm going to want to change the prompts.
我想要修改提示詞。
6:22.360–6:22.480
Okay.
好的。
6:22.560–6:23.200
So it said,
所以它說,
6:23.740–6:24.320
TLDR,
簡短摘要,
6:24.320–6:26.700
it's a three-way AI directors comparison flow,
這是一個三方 AI 導演比較流程,
6:27.160–6:29.800
one shared start frame fans out into three C dance,
一個共享的起始幀分叉成三個 C 舞蹈,
6:29.900–6:30.380
two nodes,
兩個節點,
6:30.500–6:32.340
each driven by a different model's motion prompt.
每個節點由不同模型的動作提示驅動。
6:32.780–6:32.960
All right.
好的。
6:32.960–6:33.960
So next we're going to write,
所以下一步我們要撰寫,
6:34.380–6:38.420
please spawn three different sub agents,
請生成三個不同的子代理,
6:38.940–6:40.700
one Kimi K3,
一個是 Kimi K3,
6:41.220–6:42.380
one Fable,
一個是 Fable,
6:42.380–6:45.400
and one GPT 5.6,
和一個 GPT 5.6,
6:45.540–6:49.840
and have them each look at the reference image,
並讓它們各自查看參考圖片,
6:49.980–6:53.640
then create their own motion prompt
然後創建它們自己的動作提示,
6:53.640–6:58.760
and insert it into the relevant text box,
並將其插入相關的文字方塊中,
6:58.980–6:59.360
please.
請。
6:59.780–7:01.880
So Claude code has just replied and it said,
所以 Claude Code 剛剛回覆了,它說,
7:01.980–7:02.140
done.
完成。
7:02.220–7:04.120
All three models view the actual start frame,
這三個模型都查看了實際的起始幀,
7:04.480–7:05.640
wrote their own motion prompt,
寫下了它們各自的動作提示,
7:05.640–7:07.440
and the prompts are live within the flow.
並且這些提示已即時整合到流程中。
7:07.440–7:09.260
So you can see Kimi K3,
所以你可以看到 Kimi K3,
7:09.320–7:10.180
110 words,
110 個字,
7:10.280–7:13.360
and then it kind of describes the different actions
然後它大致描述了每個模型在此採取的不同動作
7:13.360–7:15.160
and motions that each one of them took here.
和運動。
7:15.320–7:16.760
And so now we can go back into the flow
所以現在我們可以回到流程中,
7:16.760–7:19.120
and you can see that this has already changed.
你可以看到這已經改變了。
7:19.340–7:21.600
Each prompt has been automatically inserted
每個提示都已自動插入
7:21.600–7:22.900
into their prompt box here.
到這裡的提示方塊中。
7:23.100–7:25.520
And then you can see these purple lines here.
然後你可以看到這裡的紫色線條。
7:25.640–7:26.600
This is the text connector.
這是文字連接器。
7:26.740–7:29.420
So it's connecting into the C dance to here.
所以它連接到這裡的 C 舞蹈。
7:29.820–7:32.200
And then this start frame is also getting connected.
然後這個起始幀也獲得了連接。
7:32.200–7:35.460
So each of these videos starts from the same start frame.
所以這些影片中的每一個都從相同的起始幀開始。
7:35.460–7:39.380
Then the language models can take it in different directions.
然後語言模型可以將它們帶向不同的方向。
7:39.860–7:41.020
And so I've already run these models
所以我已經運行了這些模型,
7:41.020–7:42.840
and this was the output.
這是輸出結果。
7:45.840–7:47.380
And the next steps in this workflow
這個工作流程的下一步
7:47.380–7:49.060
is essentially a image compositor.
本質上是一個影像合成器。
7:49.160–7:51.840
So it's just going to use these black overlays
所以它只會使用這些黑色覆蓋層,
7:51.840–7:54.740
and then it's sending this into a video editor.
然後將其發送給影片編輯器。
7:55.000–7:57.180
And this is essentially just putting those overlays
這本質上只是將這些覆蓋層
7:57.180–7:59.660
on those C dance to clips here
放在這裡的 C 舞蹈片段上,
7:59.660–8:02.380
so that we can post this on Twitter or LinkedIn.
以便我們可以將此發布到 Twitter 或 LinkedIn。
8:02.380–8:05.840
So everything I've shown in this video so far
所以我在這部影片到目前為止所展示的一切,
8:05.840–8:06.980
I'll include it in the description
我會將其包含在描述中,
8:06.980–8:08.380
including this template here
包括這裡的模板,
8:08.380–8:11.300
which also includes a skill in it
其中還包含一個技能,
8:11.300–8:14.040
to set up those sub-agents in Claude Code.
用於在 Claude Code 中設置這些子代理。
8:14.900–8:17.360
So hopefully you guys found this tutorial useful
所以希望你們覺得這個教程有幫助,
8:17.360–8:19.960
and you're able to set up Kimi K3 for yourselves.
並且你們能夠自行設置 Kimi K3。
8:19.960–8:20.000
Let's see.
讓我們看看。