A control-room operator does not hand a new contractor a blank badge and say "figure it out." Every person who steps onto a live switching job carries a context packet: the jurisdiction they are in, the standards they must follow, the units everyone uses, and the rule that if they are not sure they stop and ask. A well-crafted system prompt does exactly the same thing for an AI model, and building that packet correctly is the skill that separates a grid professional who gets reliable AI output from one who gets confident noise.

Why System Prompts Matter for Grid Work

When you open a conversation with an AI model and type a question, you are working inside a stateless system. The model does not know your utility's territory, your applicable reliability standards, whether you measure demand in MW or MVA, or whether your jurisdiction follows NERC FAC-001 or a different planning criteria. Without that context, the model will answer using its training distribution: a mix of utility practices from dozens of US regions, international grids, academic papers, and vendor documentation. That answer may be correct on average, but "average" is dangerous when you are writing a switching procedure for a 345 kV bus.

A system prompt is a persistent instruction block that sits above your conversation. Every user message the model reads is preceded by this block, so the context is never absent. Think of it as the pre-job briefing that every worker on a live line receives before they pick up a tool. The system prompt tells the model who it is, what world it is operating in, and what behaviors are non-negotiable.

The four levers that a grid-context system prompt must set are:

• Jurisdiction lock: which NERC region, which state PUC, which ISO or RTO, which tariff. Without this, the model may cite MISO planning criteria when you are in WECC, or reference a federal tariff clause when your rate case is before a state commission.
• Unit discipline: MW vs. MVA vs. MVAr, kV vs. per unit, MWh vs. MMBtu, dollars per MWh vs. dollars per kW-month. A single unit swap in a capacity filing can change the arithmetic by a factor of 8.76.
• Standards anchor: the specific NERC reliability standards, FERC orders, state rules, and internal planning criteria you operate under. The model should cite these documents, not paraphrase them from memory.
• The cite-or-refuse rule: if the model cannot identify a specific, verifiable source for a claim, it must say so explicitly rather than fabricating a citation. This is the single most important behavioral constraint for regulated work.

A system prompt without a cite-or-refuse rule is not a safety net. It is a confidence amplifier attached to a hallucination engine.

Anatomy of a Grid System Prompt

A well-structured system prompt for grid work has five layers. Each layer narrows the model's operating envelope toward the reliable, verifiable output you need.

Layer 1: Role and Mandate

State who the model is in this conversation. Not "you are a helpful assistant" but something that specifies the professional domain. For example: "You are a power-system reliability analyst supporting planning and compliance work at a vertically integrated investor-owned utility in the Southeast. Your primary references are NERC reliability standards, SERC Reliability Corporation regional criteria, and applicable FERC orders. You do not advise on matters outside your configured scope without flagging the gap."

The role statement matters because it constrains the model's framing. A model told it is a reliability analyst will reason about N-1 contingency adequacy when reviewing a switching plan; the same model told it is a "helpful assistant" may optimize for sounding reasonable rather than being correct.

Layer 2: Jurisdiction and Standards

List the specific standards, orders, and internal criteria that govern the work. Be concrete: "Applicable NERC reliability standards include TPL-001-5 (transmission planning), FAC-001-3 (facility connection requirements), and MOD-032-1 (data requirements). Applicable FERC orders include Order 2023 (interconnection reform). Internal planning criteria: 115/138/230/345 kV per utility Planning Criteria document version 4.2."

Do not rely on the model to know which version of a standard is currently enforceable. NERC standards have version numbers for a reason. CIP-003-9, for instance, became enforceable April 1, 2026, replacing CIP-003-8; a model that cites requirements from the prior version is not useful to you. Specify the version in the system prompt.

Layer 3: Unit and Format Discipline

Specify your unit conventions explicitly. A useful pattern: "Report all demand quantities in MW (not MVA or per unit unless explicitly requested). Report all voltage levels in kV nominal. Report energy quantities in MWh for hourly intervals, GWh for annual totals. Report cost quantities in dollars per MWh for energy, dollars per kW-month for capacity. If a calculation requires a power factor assumption, state the assumption before computing." This eliminates an entire class of errors before they happen.

Format discipline applies to outputs as well. If you want the model to produce a table, say so. If you want numbered switching steps, say so. Unstructured prose answers in an operations context create transcription risk. Specify the output structure in the system prompt and reinforce it in individual prompts.

Layer 4: The Cite-or-Refuse Rule

This is the behavioral rule that separates a grid-safe prompt from a liability risk. The rule has two parts. First: when stating a regulatory requirement, planning criterion, or engineering parameter, the model must cite the specific document, section, and version. Second: when the model does not have a verifiable source, it must explicitly say "I do not have a verified source for this claim" and stop rather than generating a plausible-sounding answer.

Write this rule in your system prompt in plain, direct language: "If you cannot identify a specific, named source document for a regulatory, standards, or planning claim, do not state the claim as fact. Instead, state that you cannot verify it and recommend the user consult the primary source. Never fabricate a document title, section number, or citation."

The cite-or-refuse rule is particularly important for:

• NERC standard requirements (the model will confuse version numbers and effective dates)
• FERC order citations (Order numbers, docket numbers, effective dates)
• State PUC rule references (highly variable by jurisdiction)
• Asset-specific data (ratings, impedances, protection settings that exist only in your internal records)
• Interconnection study results (site-specific values that have no generalized analog)

Layer 5: Scope Limits and Escalation

The system prompt should also define what the model should not do, and what it should recommend instead. Examples: "Do not issue switching orders. Do not approve work permit exceptions. Do not make final protection coordination decisions. If a question requires access to live SCADA data, real-time topology, or internal asset records not present in this conversation, say so explicitly and recommend the appropriate internal source." This prevents the model from being used outside its safe operating zone and creates a documented scope boundary that matters for audits and incident reviews.

A Reusable Grid System Prompt Template

Below is a production-ready template you can copy and adapt. It covers the five layers above. Replace the bracketed items with your utility's actual values. Keep each section short and direct; the model does not need narrative, it needs instruction.

This template is intentionally terse. The model needs constraint, not context. You will expand individual prompts with the relevant data for each task. The system prompt sets the operating envelope; the individual prompt delivers the work.

Adapting the Template by Role

The template above is a base. Different energy roles need different layer-2 and layer-3 configurations. Here are three common adaptations.

Load Forecasting and Resource Planning

A forecaster using AI to draft day-ahead narratives or IRP scenario summaries needs the system prompt to specify: the applicable IRP filing requirement (state, RTO, or FERC), the planning horizon (short-run vs. 20-year), the forecast inputs the model should treat as authoritative (weather-normalized actuals, census projections, the interconnection queue), and a specific note about step-load disclosure. Given that peak demand is forecast to grow roughly 166 GW over five years with approximately 90 GW from data centers, any forecast narrative that omits step-load uncertainty is incomplete. Build that disclosure requirement into the system prompt: "If a demand forecast is discussed, note whether step-load events from large industrial or data-center interconnections are reflected in the forecast and identify any uncertainty band."

Compliance and Regulatory Drafting

A compliance lead drafting NERC evidence narratives or self-certifications needs the most restrictive cite-or-refuse rule of any energy role. The evidence narrative that lands in the NERC compliance binder must accurately describe what actually happened, against the requirements that actually apply. The system prompt should specify: the relevant CIP and non-CIP standards by version, the enforcement effective dates (CIP-003-9: April 1, 2026; CIP-012-2: real-time data protection between control centers), and a hard rule that the model must flag any claim about a standard requirement for human verification before the draft is finalized. An AI that confidently describes a requirement incorrectly is worse than no AI at all: it writes a plausible, wrong narrative that a busy compliance analyst may not catch.

Interconnection and Study Work

An engineer using AI to accelerate study narrative drafting or intake completeness checking needs the system prompt to reference: the applicable study process (FERC Order 2023 cluster study procedures or the utility's tariff provisions), the specific facility connection requirements (FAC-001-3 and FAC-002-3 for transmission studies), and an asset-ID discipline rule. The rule: any asset identifier (substation name, line ID, generator ID, bus number) present in AI output must be verified against the utility's GIS or model-of-record before it appears in a filed document. Invented asset IDs are one of the most common hallucination patterns in interconnection work and one of the hardest to catch without a formal check.

System Prompt Failures and What They Look Like

Understanding what goes wrong when a system prompt is missing or weak is as important as knowing how to write one correctly. Three failure patterns are common in grid contexts.

The jurisdiction drift failure. A planning analyst asks the AI to summarize voltage support requirements for a proposed 230 kV line. The system prompt specifies no region. The model answers with a mix of NERC TPL-001-5 requirements (correct) and WECC regional criteria (incorrect for a Southeast utility), because WECC is heavily represented in its training data. The analyst, not deeply familiar with WECC criteria, incorporates the incorrect requirement into a draft planning report. It surfaces in a peer review three weeks later. A jurisdiction lock in the system prompt would have prevented this.

The unit flip failure. A rate-case analyst asks the AI to calculate the revenue requirement for a proposed demand-response program. The model returns a cost figure that is off by a factor of roughly 1,000 because it used kW where the analyst expected MW. No unit convention was specified. The error is caught during testimony preparation, but only after two rounds of rework. A unit discipline layer in the system prompt would have caught this at the first output.

The fabricated citation failure. A compliance analyst asks the AI to describe the requirements for CIP-003-9 low-impact BES Cyber System controls. The model provides a detailed and plausible description that includes a section reference that does not exist in the actual standard. The analyst, pressed for time, includes the reference in a draft evidence narrative. An outside auditor flags it during a pre-audit review. This failure is not rare; it is the default behavior of a model without a cite-or-refuse rule. The rule costs nothing to add to a system prompt and can save a significant enforcement exposure.

Maintaining and Versioning System Prompts

A system prompt is a living document. Standards evolve. Your planning criteria change. New FERC orders become effective. The NERC Computational Load Entity (CLE) registry, committed for delivery by December 31, 2026, will create new registration and compliance obligations that need to appear in the prompts of anyone working on large-load interconnection or reliability compliance. When a standard version changes, the system prompt must change with it.

Treat your system prompts like you treat your planning criteria documents: version-controlled, dated, reviewed annually (or whenever a major standard becomes enforceable), and accessible to the team that uses them. A simple naming convention works: grid-analyst-system-prompt-v2.3-2026-04-01.txt. Keep prior versions. If an AI-assisted document was produced with a particular prompt version, the audit trail should show which version was active.

A utility that builds a library of role-specific, jurisdiction-specific, version-dated system prompts has built institutional knowledge into its AI usage. That is exactly what the Great Crew Change demands: as experienced planners and compliance leads retire, the system prompt captures the contextual knowledge they carried in their heads and makes it reusable by the next generation of analysts.

The system prompt is how your institutional knowledge survives a staffing transition. Write it as if you are handing it to someone who just joined the team and needs to know what matters here.

Worked Example: Before and After a System Prompt

Watch what happens to the same question with and without a grid-context system prompt.

The question: "What are the planning reserve margin requirements I need to meet for the upcoming IRP filing?"

Without a system prompt: The model produces a thoughtful, comprehensive answer covering PJM's capacity-market reserve requirements, MISO's planning reserve margin targets, and general NERC guidance on resource adequacy. If your utility is in SPP and your IRP is filed with a state commission that has its own reserve margin rule, this answer is largely irrelevant and contains no verifiable citation.

With a jurisdiction-locked system prompt (SPP region, state commission IRP requirement): The model identifies that it needs to reference the state commission's IRP rules and SPP's resource adequacy framework. It states what it can describe from its training data, flags the specific documents you should verify against (the commission's IRP rules docket, SPP's criteria), and notes the effective date of the applicable requirement. If the commission updated its reserve margin rule in the past 18 months, the model will not know the updated rule, but it will flag that its training has a cutoff and recommend verification. That honest uncertainty is infinitely more useful than a confident wrong answer.

The difference is not the model's capability. It is the system prompt's constraint. The constrained model knows what it does not know. The unconstrained model does not.

Key Takeaways

• A system prompt is the pre-job briefing for an AI model: it locks jurisdiction, units, standards, and behavioral rules before any work begins.
• The four essential levers are jurisdiction lock, unit discipline, standards anchor, and the cite-or-refuse rule. Missing any one of them creates a predictable failure mode.
• The cite-or-refuse rule is the most important behavioral constraint for regulated work: the model must name a source or admit it cannot, never fabricate a citation.
• System prompts should be role-specific: a load forecaster, a compliance analyst, and an interconnection engineer each need a different standards layer and a different output-format specification.
• Version-control your system prompts the same way you version-control planning criteria documents: dated, archived, tied to the standard versions that were in effect when AI-assisted work was produced.
• Prompt failures follow predictable patterns: jurisdiction drift, unit flips, and fabricated citations. A well-structured system prompt prevents all three.
• As experienced staff retire in the Great Crew Change, the system prompt library becomes a repository of institutional knowledge in a form that new analysts can use immediately.