02 / AI · Configuration & integration
A specialist helper, not a substitute for judgment
Configuring a Salesforce-focused assistant around a narrower brief, inspectable work, and a human review boundary.
- Status
- Personal specialist environment
- My contribution
- Configured and integrated third-party tools with AI assistance
- 01Scope the question
- 02Consult the sources
- 03Inspect the proposal
- 04Approve the next action
What must a person check before this suggestion becomes a change?
A narrower job for a general-purpose tool
FDE Buddy is my experiment in configuring a Salesforce-focused AI assistant. The workbench is where project tools and files come together. Buddy is a separate assistant environment intended to support that work. They are related, but they are not the same thing.
What is mine, and what is not
Hermes, from Nous Research, supplies the underlying agent software. Models and tool integrations supply other capabilities. My part is direction, specialization, environment setup, and iteration on instructions and workflow, with AI assistance.
I chose to keep the specialist environment separate rather than treating every personal and work-related task as one undifferentiated conversation. Giving an assistant a narrower job makes it easier to ask whether its proposal is relevant. It does not make the proposal correct.
The review boundary
I want to be able to inspect the sources and proposed change before anything consequential happens. An assistant can help organize a question, consult material, or propose a next step. A person still needs to check the target, access, assumptions, and result.
The sketch above explains that division of responsibility. It is not a trace of a successful autonomous run. This page does not claim that Buddy has passed an accuracy evaluation or can replace an engineer's review.
Before accepting an agent's suggestion
A reusable review aid for this kind of experiment. It describes what to check, not a guarantee that an agent enforces every check.
Can I trace the claim back to a source?
Ask which document, file, or observation supports the claim. If a source is missing, label the statement as an assumption rather than letting confident wording carry it.
Is this the right environment and scope?
Check the intended target and the permission to act. A correct command aimed at the wrong environment is still the wrong action.
What would show that this actually worked?
Agree on a small observable check before the change. Read the returned result. A generated instruction or a success message is not a substitute for exercising the result.
Can the change be undone?
Preserve the previous state, identify the rollback, and stop when the required access or evidence is unavailable.
Still figuring out
This remains a personal specialist experiment. Iterating instructions is not the same as training a model, and a separate workspace is not an independent security assessment. I am interested in making the assistance useful without allowing its confidence to stand in for judgment.
Explore Hermes Agent, or read Clean data is a trust exercise.