It is 9:54 PM and Maria, a solo LCSW in Oakland, is staring at four lines of handwritten shorthand from a 1:00 PM session she barely remembers. The client is a 90834, CBT for major depressive disorder, recurrent, moderate (F33.1), and the note she is about to write, "patient reports continued low mood, used CBT, will continue," would not survive five seconds of payer review. This lesson teaches you to convert exactly that kind of session shorthand into a complete, audit-ready SOAP note using AI as a drafting engine, while keeping every clinical determination, every observed fact, and the final signature where they belong: with you. By the end you will have a repeatable shorthand-to-SOAP conversion workflow, a worked example you can model every note on, and the verification habit that makes an AI SOAP note for a therapist faster than typing and safer than memory.

Why Shorthand Is the Right Input, Not the Transcript

Most AI scribe marketing assumes you recorded the session. Many clinicians do not, and for good reasons: recording consent adds friction, some clients decline, some sessions touch material (custody disputes, abuse disclosures, substance use under 42 CFR Part 2) where a verbatim transcript creates more legal exposure than it resolves. The humbler input, the four to ten lines of shorthand you jot during or immediately after session, turns out to be an excellent input for AI drafting, because shorthand is already filtered through clinical judgment. You wrote down what mattered. The AI's job is expansion and formatting, not interpretation.

Think of the AI as a medical transcriptionist from the era when those were human beings down the hall. A good transcriptionist took your terse dictation and produced a formatted document. A good transcriptionist never invented a mental status exam, never decided the client's affect was "congruent," and never added an intervention you did not perform. That is the controlling analogy for this entire chapter: the AI is your transcriptionist, not your co-clinician. The moment the draft contains a clinical observation you did not supply, the transcriptionist has overstepped, and you strike it before signing.

Shorthand input also solves a problem transcripts create: signal-to-noise. A 45-minute 90834 transcript runs 6,000 to 8,000 words, most of it conversational scaffolding. AI summarizing a transcript must decide what was clinically significant, and that is a clinical decision. Your shorthand already made those decisions. You are handing the model a curated evidence list and asking only for structure, professional register, and completeness checking. That division of labor is why this workflow produces notes you can defend, and why "ai soap note therapist" searches should lead clinicians here rather than to a record-everything pitch.

The Anatomy of a 90834 SOAP Note That Holds Up

SOAP stands for Subjective, Objective, Assessment, Plan, and each letter answers a different auditor question. Subjective answers: what did the client report? It carries the client's own account, symptoms, stressors, and at least one short quotation in the client's words, marked as a quote, in past tense ("Client stated, 'I have not gotten out of bed before noon all week.'"). Objective answers: what did you observe? This is where mental status exam elements live: appearance, behavior, mood as reported, affect as observed (range, congruence, intensity), thought process, speech, plus any administered measures with actual scores, such as a PHQ-9 of 16, down from 19 three weeks ago. Objective is the section AI is most tempted to fabricate, because it consists entirely of things only a person in the room could know. If your shorthand does not contain an MSE observation, the draft must not contain one either.

Assessment answers: what does it mean clinically? It connects the diagnosis (F33.1, major depressive disorder, recurrent episode, moderate) to current presentation, names progress or lack of it against the treatment plan, and carries the medical necessity reasoning: symptoms tied to functional impairment tied to ongoing need for skilled intervention. Plan answers: what happens next? Next session date and planned focus, homework assigned, any referrals or coordination, and any changes to frequency or the treatment plan. A plan that says "continue weekly sessions" with no clinical rationale is the single most common thin-note failure in 90834 and 90837 documentation, and it is exactly the line a UnitedHealthcare high-frequency reviewer circles.

Two structural rules keep the whole thing defensible. First, every section is past tense, third person, professional register: "Client reported," "Clinician introduced," "This writer observed." Present-tense client speech ("Client says she feels hopeless") reads like a draft and drifts toward fabricated immediacy. Second, the intervention you performed appears by name in the note, with the specific technique: not "used CBT" but "introduced cognitive restructuring targeting the automatic thought 'nothing I do matters,' using a thought record completed in session." Payers reimburse skilled intervention, and skilled intervention has a name.

The Worked Example: Maria's Actual Shorthand

Here is the shorthand Maria has from the 1:00 PM session, exactly as scrawled on her notepad. Client: 41-year-old, F33.1, sixth session, CBT, 90834, 45 minutes (2:13 to 2:58 by her calendar, she started late):

"PHQ-9 16 (was 19). Sleep still bad, 5 hr. Missed 2 days work this wk, 'can't make myself care.' No SI, checked directly. Did thought record on 'nothing I do matters', found 2 counterexamples. Flat-ish affect, brightened at end. HW: thought record daily, behavioral activation walk 3x. Next wk same time."

That is 60 words. Every element of a complete SOAP note is present in embryo: a measure with a delta (PHQ-9 19 to 16), a symptom report (sleep five hours), functional impairment (two missed work days), a direct quote, a documented risk screen (suicidal ideation assessed directly, denied), a named CBT intervention with the specific cognition targeted, observed affect with change over session, homework, and a plan. Notice what Maria's shorthand contains that no AI on earth could supply: the PHQ-9 numbers, the fact that she asked about suicidal ideation directly, the affect observation, and the session minutes. These verifiable details are the spine of the note. AI can dress the skeleton; it cannot grow the bones.

The full SOAP expansion runs roughly like this. Subjective: "Client reported persistent depressed mood and low motivation, stating 'I can't make myself care.' She reported sleeping approximately 5 hours per night and missing two days of work this week due to inability to initiate her morning routine. Client denied suicidal ideation, intent, or plan when asked directly by this clinician." Objective: "Client presented on time, casually dressed and adequately groomed. Affect was constricted for most of the session with brightening observed during the thought-record exercise near session end. Mood reported as 'flat.' Speech normal in rate and volume. Thought process linear and goal-directed. PHQ-9 administered: score 16, decreased from 19 three weeks prior, consistent with moderate depression with early treatment response." Assessment: "Client continues to meet criteria for major depressive disorder, recurrent episode, moderate (F33.1). Symptoms of anhedonia, hypersomnia disturbance, and low motivation continue to impair occupational functioning, evidenced by two missed workdays this week. The PHQ-9 decrease from 19 to 16 indicates partial response to CBT; continued weekly individual psychotherapy remains medically necessary to consolidate cognitive restructuring skills and restore occupational functioning." Plan: "Continue weekly 45-minute individual CBT (90834). Homework assigned: daily thought record and behavioral activation (walking, 3x weekly). Next session scheduled for [date], focus on behavioral activation review and continued restructuring of self-efficacy cognitions. PHQ-9 to be readministered in two weeks."

AI can dress the skeleton of a note; only the clinician in the room can grow the bones. The shorthand carries the clinical facts, the model carries the formatting, and the signature carries the liability.

The Prompt That Produces This Draft

The conversion prompt has five parts, and the constraints section does most of the safety work. Here is the actual prompt text, ready to adapt:

"You are a clinical documentation assistant for a licensed outpatient psychotherapist. Task: convert the session shorthand below into a SOAP progress note. Context: 45-minute individual psychotherapy session (CPT 90834), CBT modality, adult client, diagnosis F33.1 major depressive disorder, recurrent, moderate, session 6 of an active treatment plan. Format: four labeled sections, Subjective, Objective, Assessment, Plan; past tense; third person; professional clinical register; client referred to as 'Client.' Constraints: use ONLY facts present in the shorthand. Do not add any mental status observation, symptom, score, quote, risk statement, or intervention that is not in the shorthand. If a standard SOAP element is missing from the shorthand, insert the placeholder [CLINICIAN: VERIFY OR ADD] rather than inventing content. Do not include any identifying information. Shorthand follows: [paste shorthand]."

Walk through why each constraint earns its place. "Use ONLY facts present in the shorthand" is the anti-fabrication clause; without it, models reliably invent a tidy MSE because every training-set SOAP note has one. The placeholder instruction converts the model's completeness instinct from a hazard into a feature: instead of hallucinating a missing risk assessment, it flags the gap, and a flagged gap at 9:54 PM is a gift, because it tells you what to recall and add while the session is hours old instead of weeks old. The "no identifying information" line matters because your shorthand should already be de-identified before it touches any tool without a Business Associate Agreement; the prompt is a second fence, not the first. The format clause kills present-tense drift before it starts.

Run the prompt and the draft comes back in under fifteen seconds. Now comes the part that distinguishes a clinician from a clerk: the verification pass, line by line, draft against shorthand.

Verifying Each Line Against the Shorthand Before Signing

Print the metaphorical two columns: shorthand on the left, draft on the right. Every factual claim in the draft must trace to a specific phrase in the shorthand, and every clinically load-bearing item in the shorthand must appear in the draft. Read the Subjective section first: is the quote verbatim from your shorthand, in quotation marks, attributed past tense? Models love to "improve" quotes; an improved quote is a fabricated quote, and a fabricated client quote in a signed record is the kind of thing that surfaces in board complaint discovery. Then Objective: confirm the PHQ-9 numbers digit by digit (transposition from 16 to 19, or 19 to 16, flips the clinical story from improving to deteriorating), confirm every MSE element traces to your observation, and strike anything observed by nobody. If the model wrote "thought content unremarkable" and your shorthand is silent on thought content, that line goes, or you add it from memory as your own observation, deliberately.

In Assessment, check two things: that the diagnosis code is exactly what is in the chart (F33.1, not F32.1, not F33.0; AI swaps adjacent codes with serene confidence), and that the medical-necessity reasoning is grounded in the documented facts, not inflated. If the draft escalated "missed two days of work" into "unable to maintain employment," it has overstated impairment, which feels payer-friendly until an auditor compares notes across sessions and reads the inflation as a pattern. Inflated necessity is as dangerous as thin necessity; both invite recoupment, one invites a fraud referral. In Plan, confirm the homework matches what you actually assigned and that the next-session focus is something you actually intend.

Finally, the elements only you can supply, because this is a payer-facing document: session start and stop or total minutes consistent with the 90834 time band, the date of service, and the modality named as delivered. If Maria ever bills a 90837, the 53-plus minutes of that session are a fact from her calendar and her room, never from the model. The whole verification pass on a note this length takes three to five minutes once practiced. Compare that to seventeen minutes of cold drafting at 9:54 PM, and the arithmetic across eight clients a day is the difference between finishing at 10:20 PM and finishing before dinner.

Failure Modes Specific to Shorthand Conversion

Shorthand-to-SOAP has its own characteristic failures, distinct from transcript summarization, and you should know them by name. First, abbreviation misexpansion: "BA" in your shorthand meant behavioral activation; the model may expand it as "bipolar affective," and now your note implies a diagnosis nobody made. Keep a personal abbreviation key and paste it into the prompt context if your shorthand is dense. Second, quantity hallucination: shorthand says "sleep bad" and the draft says "client reported sleeping 3-4 hours nightly." The number is invented. Any number in the draft that is not in the shorthand is presumptively fabricated. Third, intervention inflation: you wrote "thought record" and the draft narrates a full cognitive-restructuring protocol with Socratic questioning and downward-arrow technique you did not use; generic CBT language that was not used in session is one of the ten failure modes you learned to spot in the previous lesson, and shorthand conversion triggers it constantly.

Fourth, the missing-risk-assessment trap, which cuts the other way. If your shorthand omits a risk screen, the placeholder instruction makes the model write [CLINICIAN: VERIFY OR ADD] under risk, and you must resolve it truthfully. If you assessed and forgot to jot it, add it from memory, as Maria's "No SI, checked directly" shows the habit done right. If you did not assess, the note cannot say you did, full stop; documenting an assessment that never happened is fabricating a clinical act, the gravest sin a progress note can commit. The AI never scores risk, never assigns a risk level, and never decides whether your screen was adequate; it only formats what you determined. Fifth, tense and voice drift in long drafts: the model starts past tense and slides into present by the Plan section. One regenerate with the format clause repeated usually fixes it; otherwise edit by hand.

The Time Math and the Habit Loop

Be honest about what this workflow costs and saves, because the habit only sticks if the arithmetic is real. Writing good shorthand during or within five minutes after session: about two minutes, and most clinicians already do a version of it. Running the prompt: under one minute including paste and read-back. The line-by-line verification: three to five minutes for a 90834-length note. Total: roughly six to eight minutes per note at first, trending toward the 4 to 7 minutes per note that most clinicians hit once the verification protocol from the final lesson of this chapter is internalized. Against a 15-to-20-minute cold draft, across Maria's eight daily clients, that is 80 to 100 minutes returned per day, which is not a productivity statistic, it is her son's algebra homework at 7:30 PM.

The habit that makes it durable is shorthand discipline: capture, every session, the six items the AI cannot supply. One, any measure administered and its score. Two, the functional impairment specifics (missed work, missed classes, avoided activities). Three, one verbatim client quote. Four, the risk screen and its result. Five, the named intervention and its specific target. Six, observed affect or any MSE element that was clinically notable. Sixty words containing those six items beats six hundred words missing them. Your shorthand template is the upstream half of the workflow; the prompt is the midstream; the line-by-line check is downstream. And the cardinal rule sits over all of it: the clinician signs the note. The signature is a legal attestation that every word is true and that you rendered the documented service, not a formatting step. Read every word before you sign, every time, no exceptions, even at 9:54 PM. Especially at 9:54 PM.

The Applied Problem: Your Shorthand-to-SOAP Conversion Workflow

Your artifact for this lesson is a one-page Shorthand-to-SOAP Conversion Workflow, a document you will actually tape next to your monitor. Build it in four blocks. Block one is your shorthand capture template: the six mandatory items (measure and score, functional impairment specifics, one verbatim quote, risk screen and result, named intervention with target, notable MSE observation), plus your personal abbreviation key with at least ten entries spelled out (BA = behavioral activation, TR = thought record, SI = suicidal ideation, and so on). Test the key by handing your shorthand to the model with and without it and watching what "BA" becomes.

Block two is the conversion prompt itself, copied verbatim from this lesson and customized: your typical CPT codes, your modality list, your placeholder phrase. Save it where you can paste it in five seconds (a text expander, a pinned note, a saved prompt in your tool). If you use a HIPAA-compliant scribe with a BAA such as Mentalyc, Upheal, or SimplePractice's built-in assistant, adapt the constraints into the tool's custom instructions; the anti-fabrication clause and the placeholder instruction belong in every tool you touch.

Block three is the verification checklist in trace order: quote verbatim? numbers digit-checked? every MSE line traceable? diagnosis code exact? necessity grounded, not inflated? homework accurate? minutes, date, and modality supplied by you? placeholders all resolved? Block four is your time log: for your next ten notes, record minutes spent. Done looks like this: ten consecutive notes drafted from shorthand, zero fabricated facts surviving to signature, every placeholder resolved by clinician memory or honest omission, and your average note time at or under eight minutes and falling. Run the workflow on a fabricated practice case first (never paste real PHI into a tool without a BAA), then move it into your real documentation stack.

Key Takeaways

• Shorthand is a better AI input than a transcript for many clinicians because it is already filtered through clinical judgment. The model expands and formats; it does not get to decide what was clinically significant, because you already did.
• The controlling analogy is the medical transcriptionist: the AI formats your dictation but never invents an MSE, a score, a quote, or an intervention. Any factual claim in the draft that does not trace to your shorthand is presumptively fabricated and gets struck before signing.
• A defensible 90834 SOAP note for F33.1 contains the six things AI cannot supply: the actual measure delta (PHQ-9 19 to 16), specific functional impairment, a verbatim quote, the risk screen you performed, the named intervention with its specific cognitive target, and your observed MSE elements, plus the session minutes only your calendar knows.
• The conversion prompt's load-bearing clause is the constraint block: use only shorthand facts, and insert [CLINICIAN: VERIFY OR ADD] placeholders instead of inventing missing SOAP elements. A flagged gap the same evening is recoverable; a hallucinated risk assessment in a signed note is not.
• Watch the five shorthand-specific failure modes: abbreviation misexpansion, quantity hallucination, intervention inflation with generic CBT language, the missing-risk-assessment trap, and tense drift. Inflated medical necessity is as dangerous as thin necessity; one invites recoupment, the other can invite a fraud referral.
• The realistic time math is six to eight minutes per note initially, trending toward 4 to 7 minutes once the verification protocol is internalized, versus 15 to 20 minutes of cold drafting. Across eight clients a day, that is more than an hour returned, but only if the verification pass actually happens.
• The cardinal rule governs everything: the clinician signs the note, and the signature is a legal attestation that every word is true, not a formatting step. Read every word before signing, especially when you are tired, because tired is exactly when fabrications slip through.