This article is auto-synced from its in-app version in Tai.
Company admins only. The Prompt Advisor is visible to users with the Company Admin role.
The Prompt Advisor reviews how your agents perform in real use — recorded conversations, report runs, and skill content — and proposes concrete improvements you can review and apply.
When you'd use this
Your agent works, but you want it to work better — and the best evidence is what it actually did. For example: after a busy week, evaluate ten recent conversations of your customer-care agent. The Prompt Advisor points out where it dead-ended or misused a tool, and proposes a revised system prompt you can preview as a diff and apply in one click.
Steps
Evaluate an agent's conversations
In the Workbench sidebar, click Agents (under the AI Platform section) and open the agent.
Click the Advisor tab, next to Configuration. The URL updates to include
?tab=advisor, so you can bookmark or share a link straight to an agent's Advisor.Under New evaluation, pick up to 10 conversations to evaluate. The list shows conversations from everyone in your company, most recent first — click Load more conversations to page further back.
Optionally add Focus instructions, e.g. "focus on tool usage" or "check the tone with corporate clients".
Click Evaluate n conversations. You land on the run page while the evaluation runs in the background — typically a minute or two. Feel free to navigate away; the run continues.
Review findings and apply proposals
When the run finishes, it shows a summary, findings with links to the exact conversations they came from, and improvement proposals — a rewritten system prompt, skills to attach or detach, or a drafted custom skill. Configuration advice (model choice, attachment hygiene) arrives as findings to act on yourself, not as an applyable proposal.
On a prompt proposal, click Compare before/after to start a comparison run: the Prompt Advisor re-reads the evaluated conversations and predicts, turn by turn, what the proposed prompt would have changed — before you adopt anything. It is a prediction, not a re-run: the conversations are not replayed and no tools are called again.
Click Apply to accept a proposal. Applying changes the agent immediately, so review the preview first — nothing is ever applied automatically.
Ask for an improved prompt
Findings tell you what is wrong; this turns them into something you can apply.
At the bottom of a finished run, click Propose an improved prompt. The Prompt Advisor rewrites the system prompt so it addresses the findings above, and suggests any skill changes they justify.
Optionally add Focus instructions first, e.g. "focus on the data-source findings", to weight part of the analysis.
You land on a new run showing the result — a full prompt diff, plus Compare before/after and Apply, exactly as on any other proposal.
It always works from the agent's definition as it stands now, not as it stood when the evaluation ran, so a prompt you have edited since is respected. It runs as its own Prompt Advisor run, so it counts toward your daily limit.
This is also how to recover a run that produced findings but no proposals: the findings are all it needs.
Analyze a report run
In the Workbench sidebar, click Reports and open a run from the run history.
Click Analyze this run. The Prompt Advisor checks the output against what the report promises, diagnoses the cause if the run failed or came back empty, and flags wasted reasoning that slows the report down.
Review a skill
Not yet available to your company. This area is currently managed by Nezasa and will open to company users in a future release.
In the Workbench sidebar, click Skills and open the skill.
Under the skill's Instructions field, click Review skill for a critique of the skill's routing texts and content, or Improve skill for a proposed rewrite — optionally steered by Focus instructions. Apply what you like.
Track your runs
In the Workbench sidebar, click Prompt Advisor, below Settings and Documentation.
Every evaluation, report-run analysis, and comparison is listed with its status and progress. Open a run to revisit its findings, or cancel or delete it from the row menu.
Tips and limits
Prompt Advisor runs use your company's AI usage allowance, like chats and reports — see Billing. Runs are capped per day.
An evaluation covers up to 10 conversations. For a representative picture, mix routine conversations with ones that went wrong.
Built-in agents can be evaluated but not modified, and their instructions stay private. So on a built-in the Prompt Advisor describes what it found in its own words rather than quoting, and its suggestions are limited to attaching or detaching skills and flagging knowledge gaps — Propose an improved prompt is not offered there. Agent changes can only be applied to company and personal agents; use the findings to shape your own custom agents.
Focus instructions steer what the Prompt Advisor looks at; without them it applies the full quality rubric.
The Prompt Advisor proposes, you decide: no proposal is ever applied without your explicit confirmation, and prompt changes always show a diff first.
For help while writing an agent — drafting, reviewing, and improving a system prompt before it has any conversations — use the Prompt Advisor panel in the agent editor. See Custom agents.
Related