Better prompts. Clearer next steps.
Use these prompts as starting points for review—not as permission to change a system. Replace context with your own redacted evidence and check every proposed action through the normal approval process.
Name the question and the audience
Describe the decision the reader needs to make. A request to identify missing release evidence is easier to evaluate than a broad request to improve operations. Specify the output format and allow the answer to identify what it cannot establish.
Label the evidence
Separate observed behavior, recent changes, redacted logs, and relevant configuration. State the environment and timezone where they matter. Ask the model to distinguish source evidence from hypotheses and to avoid inventing command output or unseen files.
Enforce boundaries outside the prompt
A prompt can explain a read-only policy, but the actual tool configuration must enforce it. Keep secrets out of unnecessary context and review the data-handling arrangement of the tool you use. Version reusable prompts alongside representative evaluation examples.
Four prompts. A clearer brief.
Copy a template and supply your own redacted context in the tool you use. Clipboard access depends on the browser; when it is unavailable, the text is selected for manual copying.
Review the supplied public file inventory and actual link-check results. Identify missing assets, inconsistent canonical URLs, and unresolved internal links. Treat only listed files as available. Return: finding, supporting evidence, uncertainty, next validation step. Do not execute commands or claim a deployment occurred. Ask for missing evidence rather than inventing it.Summarize the supplied redacted incident notes. Preserve timestamps and the stated timezone. Separate observed events from hypotheses. For each hypothesis, identify supporting and conflicting evidence. Do not infer a root cause from timing alone. End with the least intrusive next diagnostic step and explain what it could establish. Do not perform actions.Review the supplied role-to-action inventory. Identify permissions that appear broader than the stated task, missing target boundaries, and recovery access that needs documentation. Do not assume undocumented controls exist. Return a proposed review checklist with evidence and unanswered questions. Do not grant, revoke, or modify access.Review the described AI task and proposed output schema. Identify required inputs, acceptable outputs, uncertainty cases, and actions that must remain behind deterministic authorization. Suggest representative evaluation cases, including missing data and a provider timeout. Do not claim model accuracy or savings without measured evidence. Return a proposal for human review.