LLM Playground
Chat with Skytells-hosted large language models — configure system prompts, adjust parameters, iterate on outputs, and compare model responses without writing code.
The LLM Playground is Skytells' interactive environment for large language models — text, code, reasoning, and instruction-following models. It provides a conversational chat interface with a system prompt editor, model selector, and parameter controls, giving you everything you need to iterate on prompts and evaluate model behavior before integrating into production.
Accessing the LLM Playground
In the Skytells Console, go to AI → Playgrounds → LLM Playground in the left sidebar.
Playground layout
The LLM Playground has three primary panels: the Model & Settings panel on the left, the Chat panel in the center, and the System Prompt panel on the right.
Model selector
At the top of the settings panel, a dropdown lets you choose which language model to use. Available models include Skytells-hosted LLMs as well as compatible third-party models. The model is applied to all subsequent chat turns.
When you change the model, the conversation history is retained so you can compare how different models respond to the same messages.
System prompt
The system prompt field appears in the right panel. Type instructions here that shape the model's behavior throughout the entire conversation — persona, output format, constraints, tone, and so on.
Examples:
You are a concise technical writer. Always respond in plain text with no markdown.You are a senior software engineer with expertise in TypeScript and distributed systems.
Answer technical questions precisely and always include code examples.You are a product manager assistant. Help the user write PRDs, meeting notes, and feature specs.Changes to the system prompt take effect on the next message you send. They do not retroactively rewrite existing turns in the conversation.
Chat panel
The chat panel is the main interaction area. Type a message in the input box at the bottom and press Enter (or the send button) to submit. The model's response streams in real time.
Each message turn shows:
- Your message on the right
- The model's response on the left, with a copy button
Use the Regenerate (↻) button to rerun the last message without the model's previous response, useful for exploring different outputs to the same input.
Parameter controls
Below the model selector, a set of parameter sliders and inputs let you adjust generation behavior:
| Parameter | Description | Range |
|---|---|---|
| Temperature | Controls output randomness. Higher values produce more creative and varied outputs; lower values produce more deterministic and predictable outputs. | 0 – 2 (default: 1) |
| Max tokens | Maximum number of tokens the model will generate in a single response. | Model-dependent |
| Top P | Nucleus sampling threshold. Restricts generation to tokens whose cumulative probability is within the top P. | 0 – 1 (default: 1) |
| Frequency penalty | Penalizes tokens that have already appeared in the output, reducing repetition. | –2 – 2 (default: 0) |
| Presence penalty | Penalizes tokens that have appeared anywhere in the context, encouraging the model to explore new topics. | –2 – 2 (default: 0) |
For most use cases, start with defaults. Increase Temperature (0.7–1.2) for creative tasks and decrease it (0.0–0.3) for factual, code generation, or structured output tasks.
Working with the LLM Playground
Testing a prompt from scratch
Select a model
Open the model dropdown at the top of the settings panel and select the language model you want to test. If you are unsure which to start with, use the default — typically the latest GPT or Skytells instruction-following model.
Write a system prompt (optional)
In the system prompt panel on the right, type the persona or behavioral instructions for the model. Leave blank to use the model's default behavior.
Type your first message
Click the chat input at the bottom of the chat panel and type your message. Press Enter to send.
Review the response
The response streams in. If it is not what you expected:
- Adjust the system prompt to be more specific.
- Reduce Temperature for more focused responses.
- Click Regenerate to get a different output for the same message.
Iterate
Continue the conversation, refine your prompt, or switch models to compare outputs.
Comparing models
To compare the same conversation across multiple models:
- Complete a conversation with one model.
- Open the model selector and choose a different model.
- The conversation history is retained. Send the same message again to see how the new model responds.
- You can also clear the conversation and paste the same prompt fresh.
Getting production code
Once you've validated your prompt and parameters, implement your integration using the Inference API and the TypeScript or Python SDK.
The Skytells Inference API is OpenAI-compatible — any client built for the OpenAI API works with a Skytells API key.
Clearing and resetting
- Clear conversation — Removes all messages from the current session and starts fresh. Does not change the model or parameters.
- Reset parameters — Returns all sliders (temperature, top-p, etc.) to their defaults for the selected model.
- Change system prompt — Takes effect on the next message sent. No need to clear the conversation.
LLM Playground vs. Inference Playground
| Feature | Inference Playground | LLM Playground |
|---|---|---|
| Model types | Image, video, audio, music | Text, code, reasoning, multimodal LLMs |
| Interface | Schema-driven form | Chat interface |
| System prompt | — | Yes |
| Output | Image/Video/Audio preview | Text response |
| Code generation | Node.js, Python, HTTP snippets | Manual from API reference |
| Prediction tracking | Yes, in Predictions history | No (uses Inference API, not Predictions API) |
| Streaming | Not applicable | Yes, responses stream in real time |
Related
- Playground Hub — navigate between the two playground environments.
- Inference Playground — schema-driven form for image, video, audio models.
- Model Catalog — browse all available LLMs in the text/code filter.
- API: Inference (Chat Completions) — full API reference for the OpenAI-compatible endpoint.
- Eve — ask Eve to help you navigate to a model or explain a parameter.
How is this guide?
Inference Playground
Test image, video, and audio generation models interactively with a schema-driven form, live preview, JSON output, and automatically generated SDK code snippets.
Eve
Eve by Skytells is a master-class AI agent — orchestrator, creator, and operator. She coordinates multi-agent systems in real time, controls a live browser, executes account actions, and generates content on demand, all from a single conversation inside the Console.