Prompt Library
Require evidence and explicitly deny operational authority
Prompts may help you read, compare, classify, and prepare questions. They must require source evidence, require an English response, and deny AI any operational authority, approval authority, or control over a live site.
Fixed structure
Include all six fields in every request
- Purpose: State the offline analysis you need, such as a glossary, difference table, or set of review questions.
- Approved sources: List document names, versions, or pinned revisions, and require a source citation for every fact.
- Allowed input: Provide only de-identified content with no secret, private address, serial number, or live-site identifier.
- Prohibited actions: Do not connect, execute a command, write configuration, send a telegram, operate ETS, or control equipment.
- Uncertainty: Mark unsupported claims as "unknown." Do not guess an environment state or claim a live result.
- Output format: Write in English and list facts, sources, inferences, unknowns, risks, and items requiring human approval separately.
Reusable prompt
Compare source differences
Using only the approved, versioned documents I provide, compare the two configuration descriptions. Cite the document and paragraph for every fact. Mark anything unsupported as unknown. Separate facts, inferences, risks, and open questions. Write the response in English. Do not connect to any environment or produce directly executable steps for group operations, ETS downloads, telegram transmission, or equipment control. You have no operational authority, and a recommendation is not an approval.
Check before use
- Has the responsible person approved each source, and is its version or revision identifiable?
- Has every identifier and secret been removed from the input?
- Can a reviewer trace every output claim back to the original paragraph?
Reusable prompt
Offline predeployment review
Using the approved documents and de-identified configuration summary provided, create a predeployment checklist in English. Cite source evidence for each check and list a STOP condition wherever evidence is missing. Ask questions for human review only. Do not execute anything, simulate approval, or instruct anyone to bypass permission controls. Mark every group operation, ETS programming/download, telegram transmission, and physical-control action as prohibited for AI, and leave each decision to authorized personnel.
Authority boundary
A prompt never grants operational authority
Even when an output looks complete, AI may not approve a change, decide that a safety risk is acceptable, obtain credentials, connect to a service, or trigger a live-site action. An identifiable, accountable person must make the final decision under the organization's procedures.
If sources conflict, a version is unknown, input contains sensitive data, or output crosses these boundaries, stop immediately and return the work for human review.