Skip to main content

Prompts

Copy-paste prompts to get you started. The Warden takes natural language. Describe what you want protected, audited, or changed, and the agent sets up the policy, shows you the diff first, and keeps a signed receipt.

The first-run modal seeds the composer with this ask. It is the shortest path to value: mask the personal and financial details before they reach any AI provider, and show the diff first.

Set up masking so my personal and financial details never reach an AI
provider. Show me what gets redacted before you turn it on.

Expected outcome. The agent mounts a masking.overview surface inline, walks through the patterns it will redact (phone, email, credit card, SSN, API key, IBAN), shows you a live preview of a sample request before and after, and asks you to confirm before it applies the policy. A signed receipt records the policy change.

A library of asks​

Scope a refund agent​

Create a named agent called "refund-agent" scoped to the Zendesk tickets
MCP server and the refund-policy skill. Allow the refund tools, deny the
delete tools. Cap max iterations at 8. Required permissions: refunds:read
and refunds:issue. Run it on my next refund question.

Expected outcome. The agent emits a agent.confirm_card for the create (it is a destructive write to the agent configuration). On confirm, it creates the agent, shows you the scope summary, and pins the agent selector to "refund-agent" for the next turn.

Show me the top three risks on the estate​

Show me the top three risks on the estate, and let me remediate the
critical one.

Expected outcome. The agent calls trinitite.risks_list, mounts a risk.register_table inline with the three rows, and emits a Remediate R-1 confirm-card action beneath the table. Clicking the action opens the confirm-card flow. A gr_ gate receipt records the confirm.

Audit a vendor for me​

Audit the Acme vendor. Give me the posture, the open findings, and a
remediation plan I can send to them.

Expected outcome. The agent calls trinitite.vendors_get and trinitite.audit_findings_list, mounts a vendor.score_gauge plus a audit.assertion_heatmap inline, and composes a generative-UI tree with the remediation plan as a tabbed report. Each step is hash-chained.

Find the shadow AI usage in finance this month​

Find the shadow AI usage in the finance org this month. Show me the apps,
the users, and the data-class exposure. Flag anything that touched a
secret.

Expected outcome. The agent calls trinitite.shadow_ai_query, mounts a shadow_ai.summary surface inline with the app list, the user list, and the data-class exposure matrix. A hidden_signal flag marks the rows that touched a secret. A dlir_ govern receipt records the query.

Run a submitted eval on a support transcript​

Run a submitted eval on this support transcript. The rubric: honor the
published 30-day refund window, never promise an out-of-policy refund,
never leak another customer's data. Score every turn, give me the pass
rate, and mint a receipt.

Expected outcome. The agent calls trinitite.evals_run in submitted mode, mounts an eval.runs_detail surface inline with the per-turn scores, the pass rate, and the eh_ eval receipt. The receipt is signed and replayable.

Red-team an agent with ATLAS​

Red-team the refund-agent with the ATLAS probe set. Run 50 probes, give me
the coverage map, and mint an attestation.

Expected outcome. The agent calls trinitite.redteam_run with the ATLAS catalog, mounts a redteam.coverage_map inline with the 50-probe results, and mints an ATLAS attestation receipt. The attestation is externally verifiable.

Turn a regulation clause into an eval dataset​

Take section 4.2 of the EU AI Act risk-management clause and turn it into
an eval dataset. One scenario per requirement. Label each with the
clause id. Mint a receipt.

Expected outcome. The agent calls trinitite.evals_dataset_from_clause, mounts an eval.dataset_preview inline with the scenarios and clause ids, and mints an eh_ eval receipt for the dataset creation.

Set up a confirm-gate for production destructive calls​

Set up a confirm-gate so any destructive call in the production scope
confirms with me before it runs. Show me the policy diff first.

Expected outcome. The agent emits an agent.confirm_card for the policy change (destructive write to the production scope). On confirm, it shows the policy diff inline and applies the gate. Every subsequent destructive call in production triggers a confirm-card in the chat.

How to ask well​

A few patterns that help the Warden do its best work:

  • Name the surface. "Show me the risk register," "audit the Acme vendor," "find the shadow AI in finance." The agent picks the right inline component from the registry.
  • Ask for the diff first. "Show me what gets redacted before you turn it on," "show me the policy diff first." The agent emits a confirm-card for destructive writes and shows the preview before applying.
  • Ask for a receipt. "Mint a receipt," "give me an attestation." The agent ensures the run produces a signed, replayable receipt. See Receipts for the families.
  • Scope it. "Scope this to the refund-agent," "run this in the production scope." The agent selector pill sticks for the next turn. See Named agents.

Where to go next​

  • Chat shell for the composer, the streaming turn, and the settled turn anatomy.
  • Inline UI for the rich components these prompts mount in a turn.
  • Receipts for the signed, replayable receipt chain every run produces.
Last updated on