What is Feluda Workbench?

What is Feluda Workbench?

Feluda Workbench is the interactive area where you work directly with an AI model.

You can use it to ask questions, test instructions, compare models, attach supported content, enable approved tools, and review the actions taken during a conversation.

Workbench is a good place to begin when you are exploring a task or are not yet ready to turn it into a repeatable workflow.

What you can do in Workbench

Workbench can help with many everyday tasks.

You can use it to:

• summarise notes, documents, or messages;
• rewrite or improve text;
• extract useful details;
• classify information;
• compare ideas or options;
• prepare an outline or draft;
• test how different AI models respond;
• use tools provided by installed Genes; and
• check whether a task is suitable for a workflow.

The result depends on the selected model, the instruction you provide, and any tools you choose to enable.

Choose a provider and model

Before starting a conversation, select an AI provider and one of its available models.

The provider may be:

• a supported cloud AI service; or
• a compatible local model running on your own computer.

Different models may respond differently to the same request. They can vary in speed, writing style, reasoning ability, supported features, and cost.

You can change the selected model when you want to compare results or use a different model for another task.

If no models appear, open AI Providers and confirm that at least one provider has been connected correctly.

Start a conversation

A Workbench conversation begins with an instruction.

For example:

Summarise the following notes in five bullet points.

Include:
1. the main decisions;
2. the people responsible for each action; and
3. any unanswered questions.

Notes:
[Paste the notes here.]

A clear instruction usually explains:

• what the model should do;
• which information it should use;
• what the answer should contain; and
• how the answer should be presented.

You can continue the same conversation by asking the model to shorten, clarify, reorganise, or improve its earlier response.

Workbench remembers the conversation

The selected model can use earlier messages in the same conversation as context.

This means you can work in stages.

For example, you could:

1. ask for a summary;
2. ask for missing details;
3. request a table;
4. ask for a shorter version; and
5. prepare a final draft.

Keep in mind that a very long conversation can become harder for a model to follow. Start a new conversation when the task changes or when earlier messages are no longer useful.

Use supported attachments

Workbench may allow you to add supported files or media to a conversation.

Attachments can help when the task depends on source material that would be inconvenient to paste into the message.

Depending on the selected model and your Feluda setup, supported content may include images, audio, or files.

Before attaching anything:

• confirm that the selected model can use that type of content;
• remove information that is not needed;
• check whether the model is local or cloud-based; and
• review the provider's privacy terms when using an online service.

The model may not interpret every attachment correctly. Check important details against the original source.

Use tools in Workbench

Tools let an AI model do more than prepare a written response.

A tool may allow the model to:

• retrieve information;
• work with files;
• search an approved source;
• create a Journal entry;
• use a connected service; or
• perform another action provided by an installed Gene.

Available tools depend on the Genes and connections in your Feluda setup.

To use them:

1. open Workbench;
2. select Tools;
3. review the available tools;
4. enable only the tools needed for the task; and
5. give the model a clear instruction.

Enabling a tool makes it available to the model during the conversation. It does not mean that the tool will be used for every message.

Stay in control of tools

Review a tool before enabling it.

Check:

• what the tool is designed to do;
• what information it may receive;
• whether it connects to an outside service;
• whether it needs a private setting or account; and
• what kind of action it can perform.

Start with non-sensitive information while testing a new tool.

Use only the tools required for the current task. Disable tools that are not needed.

Review the Activity log

When the model uses a tool, Workbench records the activity for the conversation.

The Activity log helps you review:

• which tool was used;
• when it was used;
• whether the action completed;
• what result was returned; and
• whether an error occurred.

Do not rely only on the model saying that an action was completed. Check the Activity log and, when possible, confirm the result in the place where the action should have happened.

For example, if the model says it created a Journal entry, open the Journal and confirm that the entry exists.

Compare AI models

Workbench is useful for comparing models before choosing one for regular work or a workflow.

Use the same instruction and source material with each model.

Compare:

• how closely the model follows the instruction;
• whether it includes the important details;
• whether it adds unsupported information;
• how clearly it presents the answer;
• how quickly it responds; and
• whether it supports the tools or attachments you need.

Use sample information during comparison so that every model receives the same input.

Review every important response

AI-generated answers can be incomplete, incorrect, or overly confident.

Before using an important response, check whether it:

• answers the actual request;
• follows the requested format;
• matches the source information;
• leaves out important details;
• introduces claims that were not provided; and
• is suitable for the intended audience.

Human review is especially important for work involving customers, money, legal matters, health, safety, security, or other sensitive decisions.

Workbench and Studio are different

Workbench is best for interactive work.

Use it when you want to explore, ask follow-up questions, test an instruction, or compare models.

Studio is best for a task with clear steps that should be repeated.

A useful way to work is:

1. test the task in Workbench;
2. improve the instruction;
3. confirm the expected result;
4. build the repeatable process in Studio; and
5. run it with new information through RunFlows.

You do not need to turn every Workbench conversation into a workflow.

A simple Workbench example

Imagine that you receive a customer message and need to understand it quickly.

You could enter:

Read the customer message below.

Return:
1. the main issue;
2. the urgency;
3. any missing information; and
4. a polite draft reply.

Customer message:
[Paste the message here.]

Review the answer, correct anything that is inaccurate, and ask a follow-up question when needed.

If you perform the same task often, you may later turn it into a workflow.

When to start a new conversation

Start a new conversation when:

• you begin a different task;
• the earlier messages are no longer relevant;
• you want to compare a clean response from another model;
• the current conversation has become confusing; or
• you are about to work with unrelated sensitive information.

A focused conversation helps the model understand what matters.

If the model does not respond

Check that:

• a provider and model are selected;
• the provider is still available;
• your internet connection works when using a cloud model;
• the local model application is running when using a local model;
• the request is not too large for the selected model; and
• the provider account can use the model.

Try a short message to confirm the basic connection.

If the short message works, reduce the size of the original request or divide the task into smaller parts.

If a tool is not available

A tool may be missing because:

• its Gene has not been added or synced;
• the required connection has not been configured;
• a private setting is missing;
• the tool is not enabled for the conversation; or
• the selected model does not support the required tool use.

Review the Gene or tool guide, complete any required setup, and reopen the Workbench tool list.

Start with one useful task

For your first Workbench session:

1. select one provider and model;
2. enter a short, clear instruction;
3. review the answer;
4. improve the instruction once; and
5. try one approved tool only when you understand what it does.

Once you can get a useful and repeatable result, continue with the guide to writing clear AI instructions or begin building a small workflow in Studio.