AI Scribe Prompts for Geriatric Cognitive Assessment: MoCA and SLUMS Documentation Playbook

Documenting MoCA and SLUMS results with an AI scribe requires structured prompts that capture the instrument version, raw score, education-adjusted score, domain subscores, administrator, and clinician interpretation. The note must then link findings to G31.84 or F03.90 and to every required CPT 99483 care planning element before signature.

AI scribe prompts must capture raw item scores, not narrative summaries

CLINICAL UPDATE 2026: Revised for 2026 CMS CPT G2211 attestation rules, California SB 1120 compliance, and NCCI Modifier 25 bundling criteria.

Ambient AI scribes default to narrative compression. A 12 minute MoCA administration becomes "patient scored in the low 20s, some memory trouble." That sentence is clinically useless, defenseless in an audit, and cannot support CPT 99483, a dementia staging decision, or a specific ICD-10-CM code. Payers and Medicare Administrative Contractors (MACs) look for the instrument name, the numeric total, the denominator, and evidence that the result changed clinical management.

The fix is prompt architecture. Your scribe configuration must instruct the model to extract discrete values into fixed fields, to flag missing fields rather than infer them, and to keep the clinician's interpretation separate from the test result. For a Clinic Medical Director, this is a governance decision, not a provider preference. One standardized cognitive prompt set across all geriatric, primary care, and neurology clinicians produces consistent notes, consistent coding, and consistent audit defense.

Scribing.io builds this structure natively. The Scribing.io Clinical Interface treats each cognitive instrument as a structured documentation object with required child fields. If the delayed recall subscore is absent from the transcript, the draft note displays an unresolved field instead of a guessed number. That single behavior eliminates the most common source of AI hallucination in cognitive documentation.

Never let the model estimate a cognitive score from conversation.

MoCA documentation requires version, education adjustment, and seven domain subscores

The Montreal Cognitive Assessment is a 30 point screening instrument. A score of 26 or above is generally considered within normal limits, and one point is added for patients with 12 or fewer years of formal education (capped at 30). The test assesses visuospatial and executive function, naming, attention, language, abstraction, delayed recall, and orientation. Since 2019, the copyright holder has required administrator training and certification for clinical use, which makes the administrator field an auditable compliance element.

A defensible MoCA entry contains the following discrete elements, each of which your prompt must request by name:

• Instrument and version used: MoCA 8.1, 8.2, or 8.3 (alternate forms reduce practice effects on repeat testing), MoCA Blind, or MoCA 5 minute version.

• Administrator name and certification: clinician, nurse, or medical assistant, with certification status recorded.

• Raw total score: written as a fraction, such as 21/30.

• Years of formal education: with the education point applied or not applied, and the adjusted total.

• Domain subscores: visuospatial/executive (/5), naming (/3), attention (/6), language (/3), abstraction (/2), delayed recall (/5), orientation (/6).

• Memory Index Score if calculated: useful for distinguishing amnestic from non-amnestic presentations.

• Confounders observed during testing: hearing or vision impairment, limited English proficiency, interpreter use, fatigue, acute illness, sedating medications.

• Prior scores for trending: date and score of the most recent prior administration.

Domain subscores carry clinical weight. A patient scoring 23/30 with 1/5 delayed recall and preserved executive function has a different differential than a patient scoring 23/30 with 1/5 visuospatial/executive and intact recall. The first pattern suggests an amnestic process consistent with Alzheimer disease pathology; the second raises vascular or Lewy body considerations. A narrative summary erases that distinction and weakens the medical decision making that supports moderate or high complexity coding.

SLUMS interpretation depends on education-stratified cutoffs the scribe must apply

The Saint Louis University Mental Status Examination is an 11 item, 30 point instrument developed within the VA system. Unlike the MoCA, SLUMS does not add a point for education. It applies separate interpretation bands depending on whether the patient completed high school. That difference is the most frequent SLUMS documentation error we see in AI generated notes.

For patients with high school education, 27 to 30 is normal, 21 to 26 suggests mild neurocognitive disorder, and 1 to 20 suggests dementia. For patients with less than high school education, 25 to 29 is normal, 20 to 24 suggests mild neurocognitive disorder, and 1 to 19 suggests dementia. A score of 22 therefore means mild neurocognitive disorder in one patient and normal cognition... no, mild neurocognitive disorder in both, while a score of 24 is mild impairment for the graduate and mild impairment for the non graduate as well. A score of 25, however, is mild impairment for the high school graduate and normal for the patient without a diploma.

Your SLUMS prompt must therefore require the education level before the model writes any interpretation. If education is not stated in the encounter, the scribe should output "education level not documented; interpretation band cannot be assigned" and leave the field for clinician completion. Scribing.io applies this as a hard dependency: the interpretation node cannot render until the education node is populated.

SLUMS score 25 is normal or impaired depending on education.

Five production prompts standardize cognitive documentation across every provider

The following prompt set reflects what we deploy for geriatric and primary care groups. Each prompt is written to be pasted into a custom instruction field or loaded as a specialty macro. Clinic leaders can adapt the language, but the structural rules (discrete fields, no inference, flag missing data) should not be loosened. Our 1,000+ Specialty Clinical Macros library contains maintained versions of each.

Prompt 1: MoCA structured extraction

Extract the MoCA administration into the following fields only: version, administrator, certification status, raw total out of 30, years of formal education, education point applied (yes or no), adjusted total, and each of the seven domain subscores with denominators. Record observed confounders. If any value is not explicitly stated in the transcript, write "NOT CAPTURED" and do not estimate. Do not reproduce copyrighted test items or stimulus content. Place clinician interpretation in a separate field labeled Interpretation.

Prompt 2: SLUMS structured extraction

Extract the SLUMS administration into fields for total score out of 30, education category (high school graduate or less than high school), and item level points where stated. Assign the interpretation band only after education category is confirmed, using the published education stratified cutoffs. If education is absent, output "Interpretation band pending education confirmation." Record hearing, vision, language, and interpreter factors.

Prompt 3: Diagnosis linkage and specificity

Do not assign a diagnosis from a screening score alone. List the screening result, functional status findings, and clinician stated impression. Suggest ICD-10-CM codes only when the clinician verbally states a diagnosis. For dementia, prompt for severity (mild, moderate, severe) and for behavioral, psychotic, mood, or anxiety disturbance. For mild cognitive impairment, prompt for stated etiology if known.

Prompt 4: CPT 99483 element completeness

Generate a 99483 element checklist and mark each element as documented, partially documented, or absent, citing the transcript segment supporting each. Elements: cognition focused history and exam, MDM complexity, functional assessment with decision making capacity, standardized staging instrument, medication reconciliation with high risk medication review, neuropsychiatric screening with standardized instrument, safety including driving, caregiver identification and needs, advance care plan, and written care plan shared with patient or caregiver.

Prompt 5: Caregiver and safety capture

Identify every caregiver mentioned by name and relationship. Document caregiver knowledge of the diagnosis, stated needs, available social supports, and willingness to take on caregiving tasks as separate statements. Document firearm access, driving status, wandering risk, stove and medication safety, and fall history. Quote the patient or caregiver when they describe a safety event.

These five prompts work together. Prompts 1 and 2 produce clean data; Prompt 3 prevents the model from overreaching into diagnosis; Prompts 4 and 5 build the care planning record that separates a screening encounter from a billable cognitive assessment and care plan service. For specialty specific variants (memory clinic, home based primary care, post acute SNF), see the Specialty Clinical Documentation Library.

CPT 99483 reimbursement depends on ten documented elements, not the screening score

CPT 99483 reimburses cognitive assessment and care planning for patients with suspected or confirmed cognitive impairment. It may be furnished in the office, home, domiciliary, or rest home setting, and CMS limits billing to once per 180 days per patient. The typical face to face time is 60 minutes. The MoCA or SLUMS score supports the cognition focused evaluation, but it is only one of ten required components.

Auditors reject 99483 claims most often for three gaps: absent standardized staging instrument (a screening tool like MoCA does not stage dementia; FAST or the Clinical Dementia Rating does), absent standardized depression or neuropsychiatric screen, and absent evidence that the written care plan was shared with the patient or caregiver. Each gap is a documentation failure that a well prompted AI scribe can surface before the note is signed.

Same day bundling rules matter. CMS does not allow 99483 on the same date of service as several codes, including office and outpatient E/M visits 99202 through 99215, 96127 brief emotional and behavioral assessment, and advance care planning codes 99497 and 99498. Because advance care planning is already an included 99483 element, billing 99497 separately on the same day is a predictable denial. Your scribe should be configured to warn when a draft charge capture contains these conflicts.

Staging instruments must be explicit. If the clinician uses the Functional Assessment Staging Tool, the note should show the FAST stage number and the functional observations supporting it. If the CDR is used, the global score and domain ratings belong in discrete fields. The 99483 completeness prompt marks this element absent whenever only a MoCA or SLUMS total appears.

MoCA screens cognition. FAST or CDR stages dementia. Document both.

ICD-10-CM specificity turns a screening score into a supportable diagnosis

A cognitive screening score is not a diagnosis. The clinician must synthesize history, functional status, examination, and testing into a diagnostic impression. Two codes dominate geriatric cognitive documentation: G31.84 (ICD-10-CM) for mild cognitive impairment of uncertain or known etiology, and F03.90 (ICD-10-CM) for unspecified dementia, unspecified severity, without behavioral disturbance, psychotic disturbance, mood disturbance, and anxiety.

G31.84 requires preserved independence. Mild cognitive impairment implies measurable decline with largely intact instrumental activities of daily living. If the note documents a MoCA of 22/30 but also documents that the daughter now manages all finances and medications, the functional record contradicts G31.84 and points toward a dementia code. AI scribes must capture IADL findings in the same note so the code and the clinical picture agree.

F03.90 is often too vague. Since the FY2023 code expansion, unspecified dementia codes carry severity (F03.A0 mild, F03.B0 moderate, F03.C0 severe) and disturbance specifiers for agitation, psychotic features, mood, and anxiety. When a clinician states severity, the scribe should prompt for the more specific code. F03.90 remains appropriate when severity is genuinely undetermined, but defaulting to it when the FAST stage is documented leaves risk adjustment accuracy and audit consistency on the table.

Etiology specific codes follow sequencing rules. When Alzheimer disease is diagnosed, the G30 category is sequenced first with the appropriate F02 manifestation code. Vascular dementia uses the F01 category. Your diagnosis linkage prompt should never suggest etiology; it should only structure what the clinician states and flag when a stated etiology lacks a matching code pair.

Modifier 25 and G2211 apply differently to AWV, E/M, and 99483

The Annual Wellness Visit requires detection of cognitive impairment through direct observation, with consideration of reports from the patient, family, and caregivers. A brief tool such as the Mini-Cog often satisfies this requirement. When that detection triggers a full MoCA or SLUMS plus diagnostic workup, the clinician may be furnishing a separately identifiable service. The documentation must show where the preventive service ends and the problem oriented work begins. Our guide on Managing Modifier 25 Clinical Revenue Integrity covers the split note technique in detail.

NCCI logic for Modifier 25 requires that the E/M be significant and separately identifiable from the other service performed that day. In cognitive care, the classic scenario is an AWV (G0438 or G0439) followed by a problem oriented E/M addressing newly identified cognitive decline. The E/M receives Modifier 25, and its documentation must contain its own history, assessment, and plan rather than reusing preventive content. AI scribes should produce two distinct note sections when the clinician signals a transition.

G2211 rewards longitudinal dementia management. The visit complexity add-on attaches to office and outpatient E/M visits when the clinician serves as the continuing focal point for a patient's care or provides ongoing care for a serious or complex condition. Progressive cognitive impairment fits that definition well. Since CY2025, CMS permits G2211 with Modifier 25 when the same day separate service is an AWV, vaccine administration, or other Medicare Part B preventive service. For 2026, documentation should state the longitudinal relationship explicitly rather than relying on payer inference.

G2211 does not attach to 99483 itself, because 99483 is not an office and outpatient E/M code. It belongs on the follow up E/M visits that implement the care plan: medication titration, behavioral symptom management, caregiver coordination, and repeat cognitive testing. Configure your scribe to propose a G2211 attestation statement on those visits, such as "I am the continuing focal point for this patient's longitudinal cognitive care," for clinician confirmation.

Structured fields close the audit gaps that narrative notes leave open

The table below maps each high risk documentation parameter to its requirement, its typical audit vulnerability, and the Scribing.io Documentation Object Model (DOM) mechanism that addresses it.

California SB 1120 makes clinician attestation non-negotiable for AI-assisted notes

California SB 1120, the Physicians Make Decisions Act, took effect January 1, 2025. It requires that health plans and disability insurers using AI in utilization review ensure a licensed physician or qualified health care provider makes medical necessity determinations. While the statute targets payers, its downstream effect on provider documentation is real: payer reviewers will scrutinize whether AI generated content in a submitted record reflects clinician judgment.

Medical Directors should treat this as a signal. If a payer's human reviewer cannot distinguish clinician interpretation from machine summary, prior authorization for neuropsychological testing, home health, or memory care placement becomes harder to defend. The prompt rule that separates "Test Result" from "Clinician Interpretation" directly supports SB 1120 era review standards.

Three attestation controls are essential. First, the clinician must review and sign every AI drafted cognitive note, with no auto signing. Second, any field the model marked "NOT CAPTURED" must be resolved or deleted before signature. Third, the note should carry a brief statement that AI assisted drafting was used and that the signing clinician verified all scores and impressions. Scribing.io logs each edit, giving compliance teams an audit trail showing human review.

Unsigned AI drafts are not medical records. Enforce clinician review.

A 90-day rollout plan aligns providers, staff, and revenue cycle

Days 1 through 30 focus on baselining. Pull a random sample of 25 cognitive encounters from the prior quarter. Score each against the table above. Most clinics find that fewer than half of MoCA entries include domain subscores and fewer than one third of 99483 claims document a staging instrument. Those numbers become your before metric and your business case.

Days 31 through 60 focus on configuration and training. Load the five prompts into your scribe environment, verify MoCA certification for every staff member who administers the test, and train medical assistants to verbalize scores clearly during administration ("delayed recall, two out of five"). Verbal score announcement is the single highest yield workflow change, because ambient capture depends on audible data.

Days 61 through 90 focus on measurement. Repeat the 25 encounter audit. Track field completion rate, 99483 element completeness, diagnosis to function consistency, and denial rate for cognitive codes. Share provider level results monthly. Clinicians respond to peer comparison data more reliably than to policy memos.

Ongoing governance requires a named owner. Assign one physician champion and one coding lead to review prompt performance quarterly, update prompts after each annual CMS Physician Fee Schedule final rule, and adjudicate edge cases such as interpreter administered testing or telehealth MoCA administration. Clinics participating in the CMS GUIDE dementia care model should also map prompt outputs to that model's care plan and caregiver support requirements.

Telehealth and repeat testing require additional prompt safeguards

Remote cognitive testing introduces confounders. Video MoCA administration is used in practice, but visuospatial items and the trail making task are harder to standardize remotely. Your prompt should capture the modality (in person, video, audio only), any items omitted or modified, and whether the total was prorated. Prorated or modified scores must be labeled as such and should not be compared directly to prior in person totals without comment.

Repeat testing needs version rotation. Practice effects inflate MoCA scores when the same version is repeated within months. Prompt the scribe to record the version and compare it against the prior administration's version. If the same version was used, the note should state it so the clinician can weigh the trend accordingly. This small field strengthens longitudinal documentation and supports G2211 on follow up visits.

Frequently asked questions about AI scribe prompts for MoCA and SLUMS documentation

What must an AI scribe capture for a defensible MoCA note?

A defensible MoCA note includes the version used, administrator and certification status, raw score out of 30, years of education, whether the education point was applied, the adjusted total, all seven domain subscores, observed confounders, and the clinician's separate interpretation. Missing values should be flagged rather than estimated.

How does SLUMS interpretation differ from MoCA scoring?

SLUMS does not add points for education. Instead, it applies separate cutoff bands for high school graduates and patients with less education. A graduate scoring 21 to 26 suggests mild neurocognitive disorder, while a non graduate's mild band is 20 to 24. Scribes must confirm education before assigning any interpretation band.

Can a MoCA score alone support CPT 99483 billing?

No, a MoCA score alone cannot support 99483. The code requires ten elements, including a standardized dementia staging instrument such as FAST or CDR, a standardized neuropsychiatric screen, functional and safety assessment, caregiver evaluation, advance care planning, and a written care plan shared with the patient or caregiver.

When should clinicians code G31.84 versus F03.90?

Use G31.84 for mild cognitive impairment when measurable decline exists but instrumental daily activities remain largely independent. Use F03.90 for unspecified dementia when functional dependence is documented but severity is undetermined. When severity is stated, prefer F03.A0, F03.B0, or F03.C0 for greater specificity and audit consistency.

Can G2211 and Modifier 25 apply to cognitive care visits?

Yes, with specific limits applied. G2211 attaches to office E/M visits managing longitudinal cognitive care, not to 99483. Since 2025, CMS allows G2211 with Modifier 25 when the separate same day service is an AWV or Part B preventive service. Documentation must explicitly state the continuing care relationship.