Generative AI: Models, Chatbots, Agents, and Tools

A model generates an output from the input available to it. A chat interface carries a conversation; an agent harness can combine that model with context and tools in an action-and-check loop. A fluent explanation alone proves neither that a file changed nor that a test ran.

Use a fictional static website called Cedar Studio. Its contact form has a name, an email address, and a message. No real customer records or working delivery service are needed. Keep a plain-text scratchpad for a requirement and an observation.

That request hides several different jobs. A chatbot can suggest clearer labels. An agent with file access might edit markup. A browser tool can inspect the result. A delivery service would send the message, which is a separate external action requiring explicit scope.

• Suggestion: “Add a visible email label.” Inspect the suggested wording, not a claimed file edit.
• Action: change contact.html only. Check the actual diff to see whether the label points at the input.
• Evidence: tab to the email field in a browser and confirm its accessible name. A screenshot alone cannot establish keyboard behavior.

1. Write one observable problem: “The email box has only a placeholder, which disappears while typing.”
2. Ask for a read-only explanation of the current markup and the smallest proposed correction.
3. Classify each proposed step as advice, file change, command, or external action. Approve only the category you intend.
4. After a permitted edit, inspect the file and repeat the original observation. Keep the before/after evidence together.

A useful completion report names the file changed, the behavior checked, and what remains untested. “The form is improved” is too broad. “Email has a persistent label; delivery was not tested” is bounded and reviewable.

Stop if a text-only answer claims to have deployed the site, if a worker requests live credentials for this markup exercise, or if the proposed change expands into a new backend. Ask for the missing evidence or a smaller task.

• Can you point to an artifact rather than a promise?
• Does the test examine the user-visible problem, not merely whether text was generated?
• Have you kept sending a real message outside this local exercise?