It was 6:40 in the evening and the caseworker still had three notes to write. She opened the agency's new AI tool, the one the supervisor had spent a meeting praising, and typed what felt natural: "Write a home visit note for the Ramirez family." The draft came back in seconds, a clean, fluent, professional-sounding note. It described a tidy kitchen, two children who "appeared well adjusted," and a mother who "engaged warmly with the worker." The problem was that the caseworker had not told the model any of that. She had not described the kitchen. She had not said anything about the children's adjustment. The model had taken a one-line request and filled the entire note with plausible, invented content, because that is what a vague prompt invites it to do. She deleted the draft. The tool was not broken. The prompt was. This lesson is about the difference, because in this work the prompt is not a convenience setting. It is the line between a note grounded in what happened and a note the model made up.

Why the Prompt Is the Control Surface

A prompt is simply the instruction and the material you give to an AI tool before it generates output. In an ordinary office, a vague prompt produces a mediocre email and nobody is harmed. In human services, where the output becomes a case note, a court report, or an eligibility determination, the prompt is the single most important control a caseworker has over whether the AI stays anchored to the facts of the case or drifts into invention. Before we go further, two terms: an LLM (large language model, the kind of AI that generates text from instructions and documents) does not look up facts the way a search engine does. It predicts the most likely next words given what you gave it. That single fact about how the technology works is why the prompt matters so much. The model will fill any gap you leave with what is statistically plausible, not with what is true.

Think of it the way you would think about briefing a brand-new intern who writes beautifully, works at impossible speed, has read millions of case notes, and has one dangerous habit: when they are not sure what you saw, they guess, and they write the guess in the same confident professional voice they use for the things they know. You would never hand that intern a one-line instruction and file what they produced without reading it. You would give them your field notes, tell them exactly what to do and what not to do, and check every factual claim. The prompt is how you brief that intern. The verification is how you catch the guesses. This lesson covers the briefing. The next lesson covers the catching.

Consider the time stakes, because they are real and they cut both ways. A caseworker carrying 22 families who writes one home visit note from scratch might spend 25 to 40 minutes on it. A good AI-assisted draft can cut the drafting time substantially, which across a week of visits is hours returned to the families on the caseload. That is the genuine prize. But a bad prompt does not save time. It produces a draft so unanchored that the worker either files something invented or throws it out and starts over, having lost the time twice. The skill of prompting is what turns the tool from a liability into the hours-back tool it is supposed to be.

The model fills every gap you leave with what is plausible, not with what is true. The prompt is how you stop leaving gaps.

Context: Give the Model the Record, Not a Topic

The first and most important prompting move is to give the model the actual material to work from, not a description of the material. The opening story failed because the caseworker gave the model a topic ("a home visit note for the Ramirez family") and no content. The model had nothing real to draw on, so it drew on the patterns of the thousands of home visit notes in its training data, which is to say it invented a generic, plausible visit.

Context means the raw inputs of the case: your field notes from the visit, the relevant entries from the case-management system (the software the agency runs casework in, often a state CCWIS, which stands for Comprehensive Child Welfare Information System), the prior note you are building on, the specific policy section that applies. The model can only work accurately with what is in front of it. A prompt that says "draft a home visit note based only on these field notes" followed by the actual notes is a fundamentally different and safer instruction than "write a home visit note."

The Shape of a Grounded Prompt

A grounded prompt for a case note has a recognizable structure. It tells the model who it is writing as, what document it is producing, what source material to use, and the hard rule that it must not go beyond that source. Here is the shape in plain terms, the kind a caseworker can keep as a template:

• Role and document: "You are helping a child welfare caseworker draft a formal home visit note for the case record."
• The source, pasted in full: "Use only the following field notes from today's visit: [paste the actual notes]."
• The grounding rule: "Include only observations that appear in these notes. Do not add any detail, observation, or characterization that is not explicitly in the notes."
• The gap rule: "If the notes do not contain enough information for a section, write 'insufficient information in notes' rather than filling it in."

That last instruction is the one most workers miss, and it is worth dwelling on. Left to itself, the model abhors a blank. If your field notes say nothing about the children's school attendance, a weak prompt produces a confident sentence about school attendance anyway, because notes like this usually have one. The gap rule turns the model's instinct around: instead of filling the blank with a guess, it flags the blank for you. A note that says "insufficient information in notes regarding the children's medical care" is honest and useful. A note that invents a sentence about medical care is a fabricated observation in a legal record. The single line in your prompt is the difference between the two.

Work the difference through a concrete case. A caseworker visits the Okafor home, where the concern on the case is the toddler's nutrition. Her field notes read: "Fridge stocked, formula present 2 cans, mom reports feeding q3-4h, toddler active and tracking, no concerns noted on weight per mom." A weak prompt of "write up the Okafor visit" might produce a paragraph describing the apartment's cleanliness, the mother's affect, and the older child's behavior, none of which the worker observed or recorded. A grounded prompt that pastes those exact field notes and forbids additions produces a note about the fridge, the formula, the feeding schedule, and the toddler's activity, with a flag that the worker did not document the home's general condition. The first note is faster to produce and dangerous to file. The second is the one that holds up.

The Discipline of "Only What Was Observed"

The heart of grounded prompting in this field is one phrase the caseworker must build into nearly every documentation prompt: only what was observed. This phrase, and the discipline behind it, separates a defensible case note from a liability. A case note records observation and, where the format calls for it, the worker's clearly labeled professional assessment. It does not record the model's inferences dressed as observations.

The distinction the model blurs, and that your prompt must enforce, is between observation, inference, and characterization. An observation is what was seen, heard, or measured: "Two children present, ages reported as 4 and 7. Refrigerator contained milk, eggs, and fresh produce." An inference is a conclusion drawn from observation: "The home appeared adequately resourced." A characterization is a loaded summary: "The home was a healthy environment for the children." The model, asked to write a case note, will happily produce all three in the same confident register, and a judge or an advocate reading the note cannot tell which sentences rest on what the worker actually saw. A strong prompt instructs the model to report only observations from the notes, to label any assessment clearly as the worker's assessment, and to avoid characterizations entirely.

Sorting Into Observation, Assessment, and Unknown

A reliable prompting pattern asks the model to sort the content into explicit buckets. The prompt instructs: "From the field notes, produce three sections. First, Observations: only what the worker directly saw, heard, or was told, with the source of each (direct observation versus parent report). Second, Worker Assessment: clearly labeled as the worker's professional judgment. Third, Information Gaps: anything a complete note would normally include that is not in these notes." The output of this prompt is harder for the model to fabricate into, because every sentence has to declare what kind of claim it is.

Notice the second discipline buried in that prompt: distinguishing direct observation from parent report. "The child has not eaten since yesterday" is very different in a case record depending on whether the worker observed it or the parent reported it. A weak prompt lets the model flatten the two into a single declarative sentence. A strong prompt forces the attribution: "Mother reported that the child had not eaten since the previous day" versus "Worker observed the child decline two offered snacks during the visit." In a fair hearing, which is the formal proceeding where a person challenges an agency decision and has the right to contest the evidence, the difference between an observed fact and a reported claim can decide the case. The prompt is where you protect that distinction.

"Only what was observed" is not a style preference. It is the instruction that keeps a model's inferences from entering a legal record disguised as a worker's eyewitness account.

Be Specific About Format, Audience, and Scope

Grounding the model in the source is the foundation. The second layer of prompting skill is telling the model precisely what to produce, because a model given an open-ended task will reach for length, embellishment, and the conventional structure of the document type, all of which create room for invention.

Specify the format. "Produce a SOAP-style note" or "use the agency's contact-note structure: purpose of contact, participants, observations, plan" gives the model a frame to fill rather than a blank page to populate. Specify the audience and the consequence. "This note will become part of the case record and may be reviewed by the court" is not a throwaway line; it shifts the model toward the conservative, observation-only register that the context demands, and it reminds the worker of the same. Specify the scope and the length. "Summarize this 40-minute visit in roughly 200 words covering the safety concern, the observations relevant to it, and the next step" prevents the model from padding a short visit into a long, detail-rich note where most of the added detail is invented.

Tell the Model What Not to Do

Negative instructions are as important as positive ones, and caseworkers tend to underuse them. The model does not know your professional norms unless you state them. Useful constraints to keep in a personal prompt template include: do not infer the cause of any condition you observe; do not assign clinical or diagnostic labels (the model is not licensed and neither, in that moment, are you acting as a diagnostician); do not characterize the family's compliance, motivation, or character; do not include any prior history unless I provide it; do not invent dates, ages, names, or quantities, and if a specific figure is needed but not in my notes, mark it as "[to verify]."

That "[to verify]" convention deserves emphasis because it turns the prompt into a partner for the verification step that comes next. When the model marks the toddler's age as "[to verify]" rather than guessing "approximately 18 months," the worker has a precise list of what to check before filing rather than a smooth draft that hides its own uncertain points. A worker carrying a caseload at 7 PM cannot reread every sentence with equal suspicion. A prompt that makes the model flag its own gaps and guesses tells the worker exactly where to spend the limited verification attention they have.

Consider the eligibility side, where format and scope discipline matter just as much. A benefits worker checking SNAP (the Supplemental Nutrition Assistance Program, federal food assistance), TANF (Temporary Assistance for Needy Families), or Medicaid eligibility should never prompt "is this family eligible for SNAP?" That invites the model to produce a confident determination from its training data, which may apply an outdated income threshold or the wrong rule. A disciplined prompt instead says: "Here is the household composition and the income figures I have entered: [paste]. List the specific eligibility factors the state SNAP policy manual requires me to check for this household, and for each, state what additional information or verification I need. Do not state a final eligibility determination." The prompt has been narrowed from "decide" to "organize the factors I must verify," which is the only role the tool should play, because the determination, with its due-process consequences for a family's food and shelter, stays with the human.

Iterate, Correct, and Keep the Record Clean

A first draft from a well-built prompt is still a first draft. Strong prompting is a short conversation, not a single command, and knowing how to steer the second and third turn is part of the skill. When the model produces a note that includes an inference you did not intend, the correction is specific and grounded: "The sentence about the home being a stable environment is a characterization not supported by my notes. Remove it. Replace with only the observations in the notes about the home's condition." Correcting toward the source keeps each turn anchored. Vague correction such as "make it better" invites the model to add more of its own material, which is the opposite of what you need.

A second iteration pattern is to ask the model to check its own draft against the source, which is useful but must be understood for what it is. A prompt of "review your draft and list every statement that is not directly supported by my field notes" can surface some of the model's own additions and is worth doing. But it is not verification. The same model that invented a detail can fail to flag it, or can confidently assert that an invented detail was supported. Self-checking is a helpful first pass that narrows the work; it never replaces the worker tracing each factual claim to the field notes, the case file, and the policy source. The boundary is firm: the model can help you find its mistakes, but it cannot be trusted to certify its own accuracy, because the same prediction process produced both the draft and the self-review.

Where the Prompt Ends and Verification Begins

Even a perfectly built prompt does not produce a fileable document. It produces a strong, grounded draft that is far easier to verify because it stayed close to the source, labeled its assessments, and flagged its gaps. That is the realistic prize of good prompting: not a note you can trust unread, but a draft that makes the unavoidable verification faster and more reliable. A worker who prompts well spends their verification time confirming a tightly grounded draft. A worker who prompts poorly spends it untangling invented content from real, which is slower and more error-prone, if they do it at all.

This is also where accountability lives. The cardinal rule of AI in this field is that AI informs and humans decide, and "the AI wrote it" is never a defense in a court, a licensing board, or an agency review. The prompt does not transfer accountability to the tool; it makes the worker's accountable judgment easier to exercise well. A caseworker who builds a grounded prompt, reads the labeled output, verifies every claim against the source, and signs the note has used the tool exactly as the field demands. A caseworker who types one vague line and files what comes back has signed their name to a document a prediction system wrote, and they own every word of it.

One final discipline closes the loop: privacy. The prompt is where sensitive data enters the tool, and PII (personally identifiable information, the data that identifies a specific person, such as names, dates of birth, and Social Security numbers) about families in a child welfare or benefits case is among the most sensitive there is. Before pasting field notes into any AI tool, the worker must know whether the tool is approved by the agency for that data, whether it retains what is pasted, and whether the agency's policy requires minimizing or de-identifying inputs. A brilliantly grounded prompt entered into an unapproved consumer tool can be a privacy breach even if every word of the output is accurate. Good prompting includes knowing where it is safe to prompt at all.

Key Takeaways

• The prompt is the caseworker's primary control over whether an AI tool stays anchored to the facts of the case. Because an LLM predicts plausible text rather than retrieving facts, it fills every gap you leave with invention, so the prompt's job is to leave no gaps.
• Give the model context, not a topic. Paste the actual field notes, the relevant case-record entries, and the specific policy section. A prompt that says "use only these notes" is fundamentally safer than one that names a subject and lets the model supply the content.
• Build "only what was observed" into nearly every documentation prompt, and force the model to separate observation from inference from characterization, and direct observation from parent report. Those distinctions can decide a fair hearing.
• Use the gap rule and the "[to verify]" convention: instruct the model to flag missing information rather than fill it, so its draft hands you a precise list of what to check instead of hiding its uncertainty in smooth prose.
• Be specific about format, audience, scope, and length, and use negative instructions: no inferred causes, no diagnostic labels, no characterizations of compliance or character, no unprovided history, no invented dates or quantities.
• On eligibility, never prompt the model to decide. Prompt it to organize the factors you must verify against the current policy manual, because the determination, with its due-process consequences, stays human.
• Iterate by correcting toward the source, not by asking the model to "make it better." Self-checking can narrow the work but never replaces human verification, because the same prediction process wrote the draft and the self-review.
• Good prompting does not produce a fileable note; it produces a grounded draft that is faster and safer to verify. Accountability stays with the worker who signs it, and privacy starts at the prompt: never paste family PII into a tool the agency has not approved for it.