Start Your First Conversation in Workbench
Workbench is the easiest place to confirm that your AI provider and model are working correctly.
In this guide, you will:
- open Workbench;
- choose a provider and model;
- send a first message;
- improve the result with a follow-up;
- review the response; and
- know what to do next.
You need at least one connected AI provider before you begin.
Before you start
Confirm that:
- Feluda is installed and open;
- at least one cloud or local AI provider has been connected;
- at least one model appears in Feluda; and
- your local model application is running when you use a local model.
If no models are available, open AI Providers and complete the provider setup first.
Step 1: Open Workbench
Select Workbench from the Feluda sidebar.
Workbench includes:
- a provider and model selector;
- a message area where you write instructions;
- a conversation area where responses appear;
- optional attachment controls;
- a Tools area for available capabilities; and
- an Activity log for actions performed during the conversation.
You only need the provider selector, model selector, and message area for your first conversation.
Step 2: Choose a provider
Select the provider you want to use.
This may be:
- a supported cloud AI provider; or
- a compatible local provider such as Ollama or LM Studio.
When using a cloud provider, make sure your internet connection is active.
When using a local provider, keep the local model application and its model service running.
Step 3: Choose a model
Select one of the models available through the chosen provider.
For your first conversation, choose a general-purpose chat model.
Do not worry about finding the perfect model immediately. The goal of this first test is to confirm that the connection works and that you can receive a useful response.
You can compare models later by sending the same instruction to each one.
Step 4: Send a simple test message
Begin with a short request that does not contain sensitive information.
For example:
Explain what an AI workflow is in three short sentences.
Send the message and wait for the response.
A successful response confirms that:
- Feluda can reach the selected provider;
- the model is available; and
- the conversation is working.
If the model does not respond, use the troubleshooting section later in this guide.
Step 5: Try a useful task
After the simple test works, try a task that resembles something you may use Feluda for.
For example:
Read the notes below.
Return:
1. a short summary;
2. the decisions that were made;
3. the action items; and
4. any unanswered questions.
Notes:
Sam will prepare the launch checklist by Friday.
Mia will confirm the final website copy.
The team has not yet chosen a date for the customer email.
This instruction tells the model:
- what source information to use;
- what task to perform; and
- how to organise the answer.
Review the response before continuing.
Step 6: Ask a follow-up question
Workbench keeps the earlier messages in the same conversation available as context.
You can continue by asking:
Turn the action items into a table with the columns Owner, Task, and
Deadline.
The model should use the notes and its earlier response to prepare the new format.
Follow-up messages are useful when you want to:
- shorten an answer;
- change the format;
- ask for missing details;
- correct a misunderstanding;
- change the tone; or
- continue the same task in stages.
Start a new conversation when you move to a different task.
Step 7: Review the answer
Check the response against the information you provided.
Ask yourself:
- Did the model follow the instruction?
- Did it include the important details?
- Did it use the requested format?
- Did it leave anything out?
- Did it add information that was not in the notes?
- Is the result clear enough for its intended use?
In the example above, the model should not invent a deadline for Mia or a date for the customer email.
Correct the instruction or ask a follow-up question when the answer is incomplete or inaccurate.
Improve an unclear result
When the response is too broad, add more detail to your instruction.
Instead of:
Summarise these notes.
Try:
Summarise these notes in no more than five bullet points.
Include decisions, owners, deadlines, and unanswered questions.
Do not add information that is not present in the notes.
Clear limits and a clear output format make it easier to review the result.
Try a different model
Different models may respond differently to the same request.
To compare them:
- copy the original instruction;
- start a new conversation;
- choose another available model;
- send the same instruction and notes; and
- compare the responses.
Look at:
- instruction following;
- accuracy;
- clarity;
- response speed; and
- whether the model supports features you may need later.
Use the model that produces the most useful result for your task.
Add an attachment when needed
Depending on your setup and selected model, Workbench may allow supported files or media to be attached to a conversation.
Before attaching a file:
- confirm that the selected model supports it;
- remove information that is not needed;
- check whether the model is local or cloud-based; and
- use a non-sensitive example while testing.
Give a clear instruction that explains what the model should do with the attachment.
For example:
Read the attached document and return a short summary, the main dates,
and a list of actions.
Always compare important details with the original file.
Leave tools disabled for the first test
Workbench can use tools supplied by installed Genes, but you do not need them for your first conversation.
Begin with a normal text request so you can confirm that the model itself is connected and responding correctly.
Explore tools after you understand:
- what the tool does;
- what information it may receive;
- whether it connects to an outside service; and
- how to review its activity.
Keeping the first test simple makes connection problems easier to identify.
Start a new conversation
Start a new conversation when:
- you begin an unrelated task;
- the earlier messages are no longer useful;
- you want a clean comparison between models;
- the conversation has become confusing; or
- you are changing to different source information.
A focused conversation gives the model less irrelevant information to consider.
If no provider or model appears
Open AI Providers and check that:
- the provider was saved;
- its connection details are correct;
- the cloud API key is active, when required;
- the local model application is running;
- at least one local model has been downloaded; and
- Feluda can load the provider's available models.
Return to Workbench after correcting the setup.
If the model does not answer
Try a short message such as:
Reply with the word Connected.
If that also fails, check:
- your internet connection for a cloud model;
- the cloud provider's account and model access;
- the local model service for a local model;
- whether the selected model is still available; and
- whether an error message appears in Workbench.
Restart the local model application when using Ollama or LM Studio and the connection has stopped responding.
If the response is slow
A slow response can be caused by:
- a large local model;
- limited computer memory;
- a long message or attachment;
- a slow internet connection;
- temporary cloud provider delays; or
- a model that takes more time to prepare detailed answers.
Try a shorter message or a smaller local model.
The first response from a local model may take longer while the model is loaded into memory.
If the answer is poor
A weak answer does not always mean that the connection is broken.
Try:
- making the task more specific;
- adding the desired output format;
- stating what the model should not do;
- dividing a large task into smaller steps;
- providing a clearer example; or
- comparing another model.
Review the guide to writing clear AI instructions for more help.
Protect sensitive information
Use sample information during your first conversation.
Before using private or confidential content:
- confirm which provider and model are selected;
- understand whether the model is local or cloud-based;
- review any enabled tools;
- remove details that are not required; and
- check the result before using it.
Information sent to a cloud model is handled by the selected provider. Read that provider's privacy terms before sending sensitive data.
Your first conversation is complete
Your first Workbench conversation is complete when:
- a provider and model are selected;
- the model responds to a simple message;
- you can ask a follow-up question; and
- you have reviewed the answer for accuracy.
You can now continue using Workbench, learn how to write clearer instructions, explore approved tools, or turn a repeated task into a workflow in Studio.