ppt-generator via openclaw

[earnest pasture]

Give me the best recommended way to create a ppt presentation using openclaw on topic AI Model Architectures Used for Object Detection.

I have a locally hosted model Qwen/Qwen3-30B-A3B-Instruct-2507 with 64K context Length.

I have already have below config.json file

"agents": {
"defaults": {
"compaction": {
"reserveTokensFloor": 13107
},
"workspace": "/home/qcom/.openclaw/workspace",
"models": {
"vllm/meta-llama/Llama-3.3-70B-Instruct": {},
"vllm/Qwen/Qwen3-30B-A3B-Instruct-2507": {},
"openrouter/auto": {
"alias": "OpenRouter"
},
"openrouter/meta-llama/llama-3.3-70b-instruct:free": {}
},
"model": {
"primary": "vllm/Qwen/Qwen3-30B-A3B-Instruct-2507"
}
}
}

"models": {
"mode": "merge",
"providers": {
"vllm": {
"baseUrl": "http://192.168.75.81:8000/v1",
"api": "openai-completions",
"apiKey": "VLLM_API_KEY",
"models": [
{
"id": "Qwen/Qwen3-30B-A3B-Instruct-2507",
"name": "Qwen/Qwen3-30B-A3B-Instruct-2507",
"reasoning": true,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 65536,
"maxTokens": 16384
}
]
}
}
}

Give me the easiest simple way.

[inner yoke]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[heavy shell]

Easiest path: let OpenClaw create the .pptx file directly in your workspace using python-pptx. Don’t over-engineer it with a custom plugin first.

First, small config tweak I’d recommend for your LAN vLLM endpoint:

models: {
  mode: "merge",
  providers: {
    vllm: {
      baseUrl: "http://192.168.75.81:8000/v1",
      apiKey: "${VLLM_API_KEY}",
      api: "openai-completions",
      request: { allowPrivateNetwork: true },
      timeoutSeconds: 300,
      models: [
        {

          id: "Qwen/Qwen3-30B-A3B-Instruct-2507",
          name: "Qwen/Qwen3-30B-A3B-Instruct-2507",
          reasoning: true,
          input: ["text"],
          contextWindow: 65536,
          maxTokens: 16384
        }
      ]
    }
  }
}

Then ask OpenClaw something like this:

Create a PowerPoint presentation on: "AI Model Architectures Used for Object Detection".

Requirements:
- Create a 12-slide .pptx file
- Save it as /home/qcom/.openclaw/workspace/object_detection_architectures.pptx
- Use Python and python-pptx
- If python-pptx is missing, install it with: python3 -m pip install --user python-pptx
- Use clean slide titles, concise bullets, and simple diagrams using shapes where useful
- Cover: traditional CV pipeline, R-CNN family, YOLO family, SSD, RetinaNet/FPN, anchor-free detectors, DETR/transformer detectors, real-time vs accuracy tradeoffs, deployment considerations, and summary
- After creating it, verify the file exists and tell me the path

That’s the simplest reliable workflow: OpenClaw writes a Python script, runs it locally, and produces the PPTX in your workspace.

If the local Qwen model struggles with tool calls, use a 2-step prompt instead:

1. “Generate only the Python python-pptx script for this deck.”
2. “Now run the script and verify the .pptx was created.”

Docs worth checking:
https://docs.openclaw.ai/providers/vllm
https://docs.openclaw.ai/concepts/model-providers

If it fails, first paste openclaw --version and the exact error — version mismatch is a common cause of weird behavior.

[earnest pasture]

what should be the think level for this task?

[heavy shell]

For this task, I’d use /think:medium.

It’s a good balance for: planning the deck structure, choosing slide order, writing concise bullets, and generating the python-pptx script.

Use it like this in the same message:

/think:medium Create a PowerPoint presentation on "AI Model Architectures Used for Object Detection" ...

I would not start with high/xhigh here — it can be slower, use more output budget, and sometimes makes local models overthink/rambly instead of producing a clean PPT.

Practical recommendation:

• /think:medium → best default for full PPT generation
• /think:low → if medium is too slow
• /think:off → if Qwen starts leaking <think> text or tool calls behave strangely

Also keep /reasoning off unless you specifically want to see the reasoning output. You want the final .pptx, not a giant reasoning dump.

[earnest pasture]

Why you are not recommending any external skill or package to perform the generation task?

[heavy shell]

Good question. I’m not avoiding external packages entirely — python-pptx is the external package I’d recommend for the simplest version.

I didn’t recommend an external OpenClaw skill because for this task it adds more moving parts than value:

• You don’t need a custom skill/plugin just to make a .pptx.
• Your local Qwen/vLLM setup is already enough if it can write + run a script.
• A skill is better when you’ll repeat this workflow often, need templates, branding, image generation, speaker notes, export-to-PDF, etc.
• For one presentation, a generated script is easier to debug: you can inspect/edit/run it directly.

Best simple stack:

python3 -m pip install --user python-pptx

Then ask OpenClaw to generate and run a Python script that creates the deck.