It is 9:54 PM and Maria, a solo LCSW in Oakland, has seven notes left from eight clients. She opens an AI tool, types "write a therapy progress note," and gets back 400 words of confident, generic filler that sounds like it was written about nobody, for no payer, by someone who has never sat in a room with a client. The problem is not the model. The problem is the prompt. By the end of this lesson you will be able to build a five-part clinical prompt (role, task, context, format, constraints), test it against a fabricated 50-minute CBT session for adjustment disorder with depressed mood (F43.21), critique what comes back line by line, and name exactly what is missing because the model never had your mental status exam. This is prompt engineering for therapists, taught the way a senior supervisor would teach it: slowly, with the actual text on the table.

Why the Prompt Is a Clinical Instrument, Not a Search Box

Most clinicians meet AI the way they met Google: type a few words, see what comes back. That habit is harmless when you are looking up a restaurant. It is expensive when the output is going to be the first draft of a legal medical record. A large language model does not retrieve your client's chart. It does not know the session happened. It predicts plausible clinical language from the patterns in your prompt, and if your prompt is thin, the model fills the gaps with the statistically average therapy note: vague affect, generic interventions, a plan that says "continue treatment." That average note is precisely the note a payer auditor flags, because it documents nothing specific enough to justify medical necessity.

Think of the prompt the way you think of a referral question. When a colleague sends you "please assess this client," you can do almost nothing useful. When they send "42-year-old presenting two months after involuntary job loss, sleep onset insomnia, passive death wish denied on direct inquiry, rule out adjustment disorder versus MDD, requesting diagnostic clarification and treatment recommendations," you can do real work before the client ever walks in. The model is the same. It is a brilliant, tireless, amnesiac resident who has read every textbook and met none of your clients. Everything it knows about this session, it knows because you typed it. That is the controlling idea of this entire chapter: the referral question determines the quality of the consult.

This also reframes the safety question. The model cannot leak what you never give it, and it cannot document what you withheld. Both halves matter. You will learn in later lessons exactly which platforms can lawfully receive identifiable session content under a BAA; in this lesson, every example uses fabricated, de-identified material, and that is the discipline you should carry into your own practice from day one. A good ai therapy note prompt contains clinical structure, not client identity.

The Five-Part Anatomy: Role, Task, Context, Format, Constraints

Every reliable clinical prompt has five parts. Memorize them as a sequence, the way you memorized the elements of an MSE, because you will use this skeleton for notes, treatment plans, prior authorization letters, and client handouts for the rest of this program.

Role tells the model what professional voice to write in. "You are an experienced licensed clinical social worker writing outpatient psychotherapy documentation" produces measurably different language than no role at all: more clinical register, fewer consumer-wellness phrases, correct use of terms like "affect" versus "mood." Role does not make the model a clinician. It selects which patterns of clinical writing the model imitates.

Task is the single verb-first instruction: draft, summarize, rewrite, list, critique. One task per prompt. Clinicians who stack four tasks into one paragraph ("write the note and also suggest interventions and also check the diagnosis") get a muddled answer to all four. The model, like a supervisee, does its best work on one clear assignment.

Context is the clinical material the draft must be built from: presenting problem, diagnosis, session content, interventions you actually used, client response, scores, plan. Context is where most prompts fail, and the next lesson is devoted entirely to it. The rule for now: the model may only use facts that appear in your context block, because anything else it writes is invention.

Format names the structure of the output: SOAP, DAP, BIRP, GIRP, a numbered list, a one-paragraph summary, a table. Behavioral health runs on note formats; if you do not name one, the model picks one for you, usually a hybrid that satisfies no supervisor and no Medicaid manual.

Constraints are the guardrails: length, tense, what to exclude, and the single most important instruction in clinical prompting: "Use only the information provided above. Do not add symptoms, quotes, history, interventions, or risk findings that are not stated. If something required for the format is missing, write [CLINICIAN TO COMPLETE] instead of inventing it." That one constraint converts the model from a confident fabricator into a structured drafting assistant that shows you its gaps.

The model is a brilliant amnesiac resident: it has read every textbook and met none of your clients. The prompt is the only chart it will ever see.

The Worked Example: A Fabricated 50-Minute CBT Session, F43.21

Here is the fabricated case we will use, fully invented and de-identified. The client, "R.," is a 38-year-old who lost a long-held logistics job ten weeks ago. Symptoms began within three weeks of the layoff: low mood most days, tearfulness, withdrawal from a weekly basketball game, sleep onset insomnia, guilt about "letting the family down." No prior depressive episodes, no anhedonia across all domains, symptoms clearly tethered to the stressor. Diagnosis: adjustment disorder with depressed mood, F43.21. This is the fourth session of a planned 10 to 12 session CBT course. Today's 50-minute session focused on cognitive restructuring of the automatic thought "I am useless without a job title," using a thought record. R. generated the alternative thought "my value to my family did not come from my badge" and rated belief in the original thought as dropping from 90 to 60 during the exercise. PHQ-9 administered at session start: 11, down from 14 at intake. Homework assigned: complete one thought record after each job application this week, and attend the basketball game once. Risk: R. denied suicidal ideation on direct inquiry; you assessed and documented this yourself. Plan: continue weekly CBT, review homework next session.

Now watch what happens with the search-box habit. Maria's first prompt, the one most clinicians type at 9:54 PM, is: "Write a progress note for a therapy session with a depressed client who lost his job." Here is a representative excerpt of what comes back:

"Client presented with symptoms of major depressive disorder following recent job loss. Client reported feelings of hopelessness and worthlessness. Therapist provided supportive counseling and explored the client's childhood relationship with achievement. Client stated, 'I feel like my whole world is collapsing.' Client appears to be at low risk. Will continue to monitor. Diagnosis: F32.1, Major Depressive Disorder, moderate."

Read that as an auditor would. The model upgraded an adjustment disorder to major depressive disorder and assigned F32.1, a diagnosis the chart does not support and the clinician never made. It invented a direct quote the client never said, in quotation marks, inside a legal record. It described "supportive counseling" and a childhood exploration that never happened, while the thought record, the actual billable CBT intervention, is nowhere. It rendered a risk conclusion ("low risk") with no assessment behind it. Every one of those lines is a sentence you would have to defend in a records request, and none of them came from the session. This is what a generic chatgpt for therapists notes workflow produces, and it is why the workflow, not the technology, is the hazard.

The Same Session Through a Five-Part Prompt

Now the same session through the full anatomy. This is the actual prompt text; study how each part does its job.

ROLE: You are an experienced licensed clinician writing outpatient psychotherapy progress notes that must withstand payer audit.
TASK: Draft a progress note for the session described below.
CONTEXT: Client: adult, established, session 4 of planned 10-12. Diagnosis: adjustment disorder with depressed mood (F43.21), symptoms began within 3 weeks of job loss 10 weeks ago. Session: 50 minutes, individual CBT. Focus: cognitive restructuring of the automatic thought "I am useless without a job title" using a thought record. Client generated alternative thought "my value to my family did not come from my badge"; belief in original thought dropped from 90 to 60 during the exercise. PHQ-9 today: 11 (intake: 14). Reported this week: low mood most days, sleep onset insomnia, skipped basketball game, guilt about family finances. Clinician asked directly about suicidal ideation; client denied SI, plan, and intent. Homework assigned: one thought record after each job application; attend basketball once. Plan: continue weekly CBT, review homework next session.
FORMAT: SOAP format. Subjective, Objective, Assessment, Plan as labeled sections. Past tense throughout. Refer to the client as "Client."
CONSTRAINTS: Use only the facts provided above. Do not invent quotes, symptoms, history, diagnoses, or risk findings. Do not state a risk level or interpret the SI denial; report only that the clinician inquired and the client denied SI, plan, and intent. Where the SOAP format requires information not provided, write [CLINICIAN TO COMPLETE]. Maximum 250 words.

The draft that comes back from this prompt is recognizably your session. The Subjective section carries the reported symptoms and the guilt theme. The Objective section names the thought record, the belief rating shift from 90 to 60, and the PHQ-9 of 11 against the intake 14. The Assessment ties continued symptoms to the identified stressor under F43.21 and notes measurable progress. The Plan lists the homework and the next session. And critically, where the model needed something you did not supply, it wrote the placeholder instead of writing fiction.

Critiquing the Output: What Is Missing Without Your MSE

Even the good draft is incomplete, and this is the critique skill the lesson exists to build. Read the Objective section of the five-part draft again. A defensible psychotherapy note documents the mental status examination: appearance, behavior, speech, mood as reported, affect as observed, thought process, thought content, orientation, insight, judgment. The model wrote [CLINICIAN TO COMPLETE: MSE observations] in that slot, because the MSE was never in the context. That is not a flaw in the prompt. That is the prompt working exactly as designed, and it teaches the deepest lesson in clinical AI use: the MSE is observational data that exists only in your perception of the client. No transcript contains the flattened affect you saw when R. mentioned his daughter. No prompt can transmit the half-second of psychomotor slowing you clocked when he sat down. The model can format an MSE you dictate; it can never perform one. The same is true of your risk assessment: the model can record "clinician inquired about SI; client denied SI, plan, and intent," but the inquiry, the clinical judgment about its credibility, and any risk determination are yours alone, made before the keyboard is involved.

So your critique pass on any AI draft asks three questions in order. First, is everything in the draft traceable to the context I provided, word for word where it matters (the score, the modality, the intervention)? Second, what did the format require that I never supplied, and did the model flag it or fabricate it? Third, where are the clinician-only elements, the MSE, the risk inquiry, the clinical judgment in the Assessment, and have I written those myself rather than letting the model simulate them? A draft that survives those three questions is a draft worth editing. One that fails the first question gets discarded, not repaired, because fabrication contaminates everything around it.

Run this comparison once yourself, with this fabricated case, before you ever consider AI for real documentation. Watching the same model produce a hazardous note and a serviceable one within five minutes, with the only variable being your prompt, recalibrates how much responsibility sits on your side of the keyboard. It is the single most instructive exercise in this entire level.

Why Each Part Earns Its Place: A Closer Look at Failure by Omission

It is worth seeing what each missing part costs, because in practice you will be tempted to shortcut. Drop the role, and the register slides toward wellness-blog language: "the client is on a journey of self-discovery" does not belong in a chart. Drop the task precision and ask for "help with this session," and you get a sympathetic essay rather than a note. Drop the context and you get the F32.1 disaster you already saw: invention scaled to fill the vacuum. Drop the format and you get a narrative blob your supervisor has to reverse-engineer into SOAP at review time, which costs more minutes than the AI saved. Drop the constraints and even a rich context gets embellished: the model will helpfully add a serene closing sentence about the client's "renewed hope," a clinical observation nobody made, and it will do so in fluent, confident prose that is easy to sign at 9:54 PM and hard to defend at a deposition.

There is a tense detail buried in the constraints that deserves a sentence of its own. Progress notes are written in past tense because they document what occurred; models trained on blogs and self-help content drift into present tense ("Client feels hopeful and is making progress"). Present tense reads as ongoing clinical fact rather than session observation, and tense drift mid-note is one of the tells of an unreviewed AI draft. You will meet it again in lesson three as a named failure mode. Pinning tense in your constraints prevents most of it; your read-through catches the rest.

Notice also what the five parts do to your own thinking. Writing the context block forces you to recall the session in clinical terms: what was the focus, what did I actually do, what changed, what did I measure. Clinicians consistently report that composing a disciplined context block takes two to three minutes and recovers session detail they would otherwise have lost by the seventh note of the night. The prompt is not only an instruction to the model. It is a structured recall exercise for you, and the note is better for both reasons.

The Cardinal Rule: The Clinician Signs the Note

Everything above operates under one rule that this program will repeat until it is reflexive: the clinician signs the note. Your signature on a progress note is a legal attestation that the record is accurate and that the documented services were rendered as described. It is not a formatting step, not a workflow click, not a courtesy to the EHR. When the draft contains a quote the client never said and you sign it, the fabrication becomes yours: in a payer audit, in a board complaint, in a records request from a custody attorney. No AI vendor, no matter what the marketing deck says, stands between your license and a signed inaccuracy. The five-part prompt and the three-question critique exist to make your read-before-signing fast and rigorous, never to replace it. Read every word. Every time. The day you stop is the day the seventeen-minute note from the old workflow starts looking like the safe option it never was.

This rule also settles the question pre-licensed readers ask: if you are an associate, like Carmen in Fresno, your supervisor co-signs, which means your AI draft is now their attestation too. Bringing your five-part prompt and your critique process to supervision, openly, is not a confession. It is exactly the documentation-of-judgment your supervisor needs to keep signing with confidence, and later lessons give you the supervision-agreement language for it.

The Applied Problem: Your Reusable Five-Part Clinical Prompt

Your artifact from this lesson is a Reusable Five-Part Clinical Prompt: a fill-in-the-blank template saved as a text snippet you can paste into any approved tool and complete in under three minutes. Build it now, in four steps.

Step 1: Draft the skeleton. Open a plain text file and write five labeled blocks. ROLE: "You are an experienced licensed [your license type] writing outpatient psychotherapy documentation that must withstand payer audit." TASK: "Draft a progress note for the session described below." CONTEXT: leave labeled slots: diagnosis and ICD-10 code; session number and length; modality; session focus; intervention used and client response; assessment scores with prior values; symptoms reported; risk inquiry conducted and client response; homework; plan. FORMAT: "[SOAP/DAP/BIRP/GIRP] with labeled sections, past tense, refer to the client as Client." CONSTRAINTS: copy the anti-fabrication paragraph from this lesson verbatim, including the [CLINICIAN TO COMPLETE] instruction and the prohibition on stating risk levels, then add your word limit.

Step 2: Test it on the fabricated case. Fill the slots with the F43.21 session from this lesson, exactly as given. Run it. You should get a SOAP draft with the PHQ-9 of 11, the 90-to-60 belief shift, the named thought record, and an MSE placeholder. If any of those four checkpoints fails, your context or constraints block is leaking; fix the template, not the output.

Step 3: Run the three-question critique. On the test output, verify traceability, verify the gaps were flagged rather than filled, and identify the clinician-only elements awaiting your hand. Write your answers in the margin once; this rehearses the review you will do on every future draft.

Step 4: Define done. The template is done when a colleague could pick it up cold, fill the slots from their own fabricated session in three minutes, and get a draft that contains zero facts absent from their context block. Save it where you will actually find it at 9:54 PM, and version it: this is v1, and the next two lessons will make you sharpen it.

Key Takeaways

• A clinical prompt is a referral question, not a search query. The model knows only what you type, and a thin prompt is filled with the statistically average note, which is exactly the vague, generic note that fails a medical-necessity audit.
• The five-part anatomy is role, task, context, format, constraints. Role sets the professional register, task gives one verb-first instruction, context supplies the only facts the model may use, format names the note structure (SOAP, DAP, BIRP, GIRP), and constraints forbid invention.
• The single most important constraint in any ai therapy note prompt: use only the information provided, and write [CLINICIAN TO COMPLETE] where required information is missing. This turns confident fabrication into visible, fixable gaps.
• The generic-prompt failure is predictable and serious: in the worked F43.21 example, the model invented a quote, upgraded the diagnosis to F32.1 major depressive disorder, fabricated interventions, and declared a risk level. Each of those is a sentence you would have to defend after signing.
• The MSE and the risk assessment are clinician-only. They are observational and judgment-based data that exist in no prompt and no transcript. AI can format what you observed and decided; it can never observe or decide. The model reports that you inquired about SI; it never assesses risk.
• Critique every draft with three questions: is every fact traceable to my context, did the model flag gaps instead of filling them, and where are the clinician-only elements I must write myself. A draft that fails traceability is discarded, not repaired.
• The cardinal rule governs everything: the clinician signs the note, and the signature is a legal attestation, not a formatting step. Read every word before you sign, every time, because the moment you sign a fabrication it becomes yours.