Open any activity in the Activity Inspector and read the exact system prompt your AI employee received for that run: every instruction, every tool definition, and the manifest of which skills and duties were loaded and why. Nothing paraphrased, nothing summarized, the same text the model saw.
Every AI employee runs on a system prompt assembled at the moment it starts working: instructions for its role, the tools it can call, the skills and duties that apply, and context pulled in for that specific task. Normally that assembly happens behind the scenes and you only see the result, the reply in chat or the file it produced.
See What Your Employee Is Told opens that assembly up. Inside the Activity Inspector's Prompt section, you get the full raw system prompt as text, the complete tool definitions the model was given (names, descriptions, parameter schemas), and a prompt manifest that lists every skill and duty that was included in that run, and which ones were left out. You can export the prompt to a file or copy it directly for a support ticket, an internal review, or your own audit trail.
The alternative most platforms offer is a static list of what an agent is theoretically capable of, a settings page describing skills in the abstract. This shows the actual assembled prompt for one specific run, the one that produced the answer sitting in front of you, so you are reading what really happened instead of what the product claims happens.
Because this is the most sensitive text in the whole system, the platform's most detailed configuration surfaced in plain text, it sits behind the highest plan tier and is enforced twice: the frontend hides the section, and the backend strips the prompt fields from the API response entirely before they leave the server, so there is no client-side workaround.
The raw system prompt is the literal text sent to the model for that run, not a description of it, every instruction, persona detail, and piece of injected context in the order the model received it.
Tool definitions show the exact JSON schema advertised to the model for every tool it was allowed to call in that run: names, descriptions, and parameters, the same contract the model reasoned against when it decided how to call a tool.
The prompt manifest is the audit trail behind the prompt: which skills and duties were loaded for this specific run, which were considered and left out, and the condition that made the decision, so you can see why the employee behaved the way it did instead of guessing.
A system prompt is the most complete description of how an AI employee is configured to think. Handing it out by default would expose internal prompt engineering, tool wiring, and skill logic to anyone with a login, so it is limited to the plan built for teams that need that level of control and are prepared to handle what it reveals.
The restriction is enforced in two places on purpose. The Activity Inspector hides the section in the interface, and the GraphQL layer independently strips the prompt manifest, middleware detail, and raw prompt text from the API response for any tenant without the entitlement, so the protection does not depend on the frontend alone.
The Prompt section of the Activity Inspector, gated behind the Custom (Enterprise) plan
Every activity, one complete piece of work like a chat reply or a scheduled run, is inspected the same way. Open the Activity tab, click a row or the Inspect link on a chat message, and scroll to the Prompt section near the bottom of the Activity Inspector.
On the Custom (Enterprise) plan, that section shows three things: the raw assembled system prompt as the model actually received it, the JSON schema for every tool the model was allowed to call during that run, and the prompt manifest, a tree of which skills and duties were included, which were excluded, and the condition that decided each one. A copy button and an export button sit next to the section so you can pull the text out for a ticket or a review.
On any plan below Custom, the same section renders in place but shows an upgrade prompt instead of the content. The backend enforces this independently of the screen: the GraphQL resolver checks the tenant's plan before it serializes the activity, and strips the prompt manifest, middleware details, and raw prompt text out of the response for anyone without access, so the data is never sent to a browser that shouldn't see it, not merely hidden by the interface.
When an employee's reply looks off, open the Prompt section for that exact activity and read the instructions and injected context it actually worked from, instead of guessing whether a skill misfired or the model just made a mistake.
Before trusting an AI employee with sensitive workflows, pull the raw prompt and tool definitions for a real run and review them the way you would review any system you are about to give production access to.
The prompt manifest lists every skill and duty considered for that run and the condition that included or excluded it, so you can see exactly why a capability you expected to fire did not, without re-running the task blind.
Export the exact system prompt behind a specific run and attach it to a support ticket or an internal compliance review, so the conversation is grounded in what the model was actually told rather than a paraphrase.
The Custom (Enterprise) plan. Every other section of the Activity Inspector, including Cost, Performance, and Safety, is available on every plan; the raw prompt, tool definitions, and manifest are the one part reserved for Custom.
No. Execution (Leader plan and above) shows tool calls with their inputs and outputs and the internal step trace for a run. Prompt inspection is one tier higher and shows the actual instruction text and tool schemas the model was given before it took any of those steps.
Yes. The Prompt section has an export control next to it, so you can pull the raw system prompt out for a support ticket, an internal review, or your own records without retyping it.
No. It only affects what you can view after the fact in the Activity Inspector. Nothing about how the employee runs changes, and opening the Prompt section does not consume credits.
Yes. The section still renders in the Activity Inspector on lower plans, but it shows an upgrade prompt in place of the content. The underlying data itself is stripped out by the server before it reaches your browser, not just hidden on screen.
See What Your Employee Is Told is part of How you see what they did.
A real-time activity feed shows who is working on what right now. The step-by-step inspector traces every tool call, decision, and reasoning chain. Cost tracking breaks down spend per message, per employee, and per team. Walk through your live 3D office to see your workforce at a glance.