Maria, the Oakland LCSW, hands a new client a photocopied handout on the cognitive triangle. It was written in 1997, copied through four generations of toner, and reads at college level: "maladaptive cognitions perpetuate dysphoric affect through reciprocal behavioral reinforcement." The client, a 43-year-old warehouse supervisor with major depressive disorder, nods politely, folds it into his jacket, and never reads it. That handout was supposed to extend the session into the week. Instead it said: this treatment was not written for you. By the end of this lesson you will use AI as a psychoeducation handout generator producing the same clinical content at a true 6th-grade and a true 10th-grade reading level, verify the output against CMS plain-language standards instead of trusting the model's self-report, adapt for caregivers in HCBS settings, and run it all through a repeatable workflow you can defend in supervision.

Why Reading Level Is a Clinical Variable, Not a Formatting Preference

Start with the number that should change everything a client takes home: roughly one in five American adults reads at or below a 5th-grade level, and the average adult reads comfortably at about an 8th-grade level. Health literacy runs lower still, because psychological vocabulary is a second language layered on top of reading skill. A client who can read a lease can still be defeated by "behavioral activation." When your handout outruns your client's reading level, you have not given psychoeducation; you have given homework the client cannot do and will not admit they cannot do. Nobody hands the sheet back and says "I do not understand this." They fold it into the jacket.

This is why federal health communicators treat plain language as compliance, not courtesy. The Plain Writing Act of 2010 requires federal agencies to communicate in plain language, and CMS maintains plain-language standards for materials reaching Medicare and Medicaid beneficiaries: short sentences, common words, active voice, the main point first, content organized around what the reader needs to do. States layer their own health-literacy guidance on top, and Medicaid managed-care contracts frequently specify a 6th-grade reading-level target for member-facing materials. If you serve Medicaid clients, the 6th-grade handout is not a nice-to-have; it is the version your population was always supposed to receive.

Here is the controlling analogy for this lesson: a psychoeducation handout is a dosed intervention, like a prescription. The active ingredient is the clinical concept (the cognitive triangle, sleep hygiene, the window of tolerance). The dose form is the reading level. A medication that cannot be absorbed does nothing; a handout that cannot be read does nothing. AI is very good at re-formulating the same active ingredient into different dose forms quickly. What stays your job is deciding which dose form this client needs and verifying the pharmacy did not change the active ingredient during compounding.

What AI Actually Changes About Handout Production

Before large language models, producing a custom handout meant photocopying the 1997 sheet, buying a worksheet-library subscription and hoping it had your topic at your client's level, or spending forty unpaid minutes writing it yourself at 9:54 PM, which means it never happened. Most clients received either nothing or a generic sheet pitched at whoever the original author imagined.

An LLM collapses that production cost to minutes, and it does the one thing worksheet libraries cannot: it re-renders the same concept at any reading level, in any voice you specify. The same cognitive-triangle content can exist as a 6th-grade version for the warehouse supervisor, a 10th-grade version for the verbally sophisticated graduate student who would feel condescended to by the simpler sheet, and a caregiver version for the mother of a client with schizoaffective disorder in an HCBS wraparound program. That is a real clinical gain: a client who uses the handout between sessions rehearses the skill six days a week instead of one.

But the speed creates a new failure mode you must name out loud: AI handouts look finished. The formatting is clean, the tone warm, the structure professional. The polish disguises three recurring defects. First, reading-level drift: ask for "6th grade" and you will often get 8th or 9th, because models equate "simple" with "shorter sentences" while keeping abstract vocabulary like "perspective" and "interpretation." Second, clinical drift: in simplifying, models flatten a concept into something adjacent but wrong, turning "thoughts are not facts" into "do not trust your thoughts," a different and potentially harmful message for a client with paranoid features. Third, scope creep: handout generators append advice you did not ask for, including medication suggestions or crisis instructions that do not match your actual crisis protocol. All three defects survive a casual skim. None survives the verification pass you are about to learn.

The Worked Example: One Concept, Two Doses

Let us build the actual artifact. The concept is the cognitive triangle, the CBT model showing that thoughts, feelings, and behaviors influence each other, taught to a client with major depressive disorder. Here is the prompt for the 6th-grade version; notice it specifies the audience, the reading level with a named formula, the structure, the scope limits, and a guardrail:

"You are helping a licensed therapist create a client handout. Topic: the cognitive triangle (thoughts, feelings, actions affect each other), for an adult client in treatment for depression. Write at a 6th-grade reading level measured by Flesch-Kincaid: sentences under 15 words, common everyday words, active voice, second person. Structure: a one-sentence main point at the top, a simple example using a missed phone call from a friend, one short practice exercise, a closing line telling the client to bring questions to the next session. Do not mention medication, diagnosis names, suicide, or crisis resources. Do not add content beyond the cognitive triangle. Length: under 250 words."

The 6th-grade output, after one editing pass: "Your thoughts, your feelings, and your actions are connected. Each one can change the others. Here is an example. A friend does not call you back. You think, 'She is mad at me.' That thought makes you feel sad and worried. Because you feel sad, you stay home and skip the gym. Staying home gives you more time to worry. Now try a different thought: 'Maybe she is busy.' That thought feels lighter. You might still go to the gym. Same missed call, different path. This week, catch one heavy thought. Write it down. Then write one other way to see the same event. Bring both to our next session." Count the moves: every sentence under 15 words, the concrete example before the principle, a practice exercise with exactly one step plus a write-it-down instruction.

Now the 10th-grade version of the identical active ingredient, for a more verbal client: "The cognitive triangle describes how thoughts, emotions, and behaviors continuously influence one another. When a friend does not return your call, the interpretation you reach ('she is angry with me' versus 'she is probably swamped') shapes your emotional response, and that emotion shapes what you do next, such as withdrawing or reaching out. Withdrawal then generates new evidence for the original interpretation, closing the loop. The leverage point is the interpretation: the one corner of the triangle you can examine directly. This week, when you notice a strong negative emotion, work backward to the thought that preceded it and ask what evidence supports it, what contradicts it, and what you would tell a friend who voiced the same thought." Same triangle, same missed call, same homework structure, but now the sentences carry subordinate clauses, the vocabulary includes "interpretation" and "leverage point," and the exercise has three sub-questions instead of one step. Neither version is clinically smarter. They are two dose forms of one prescription.

A psychoeducation handout is a dosed intervention: the concept is the active ingredient and the reading level is the dose form. AI can compound any dose form in minutes; only the clinician can verify the active ingredient survived the compounding.

Verifying Reading Level: Never Trust the Model's Self-Report

Here is the supervision-room rule: a model's claim that its own output is "6th grade" is a vendor claim, and you verify vendor claims independently. The tools are old, free, and everywhere. Microsoft Word displays Flesch-Kincaid Grade Level under its readability statistics; free web checkers compute Flesch-Kincaid, SMOG, and Gunning Fog from pasted text. The Flesch Reading Ease score runs 0 to 100, higher is easier; 6th-grade material generally lands around 80 or above, 10th-grade around 50 to 60. The grade-level formulas key on two things, sentence length and syllables per word, which tells you exactly how to fix output that scores too high: split sentences and swap multisyllabic abstractions ("interpretation" becomes "way of seeing it") for short concrete words.

Run the check on the actual output, every time, because drift is the rule. A first-pass "6th grade" request commonly scores at grade 8 to 9. The fix loop: paste the score back with "This scored Flesch-Kincaid grade 8.4. Revise to score below 6.5: split every sentence over 14 words and replace every word of three or more syllables with a shorter synonym unless it is a clinical term I told you to keep." Two iterations almost always converge. This loop, prompt, measure, revise, re-measure, is the heart of the workflow artifact you will build at the end of this lesson.

Then verify against the CMS plain-language standards, which go beyond what any formula measures: is the main point in the first sentence rather than buried after a warm-up paragraph? Is the voice active ("write the thought down," not "the thought should be written down")? Is the content organized around what the reader does rather than the theory's history? A handout can score Flesch-Kincaid 6.0 and still fail plain-language review because the action is buried in paragraph four. The formula checks the dose; the CMS criteria check whether the patient can open the bottle.

The Clinical Accuracy Pass: Where the License Lives

Reading-level verification is mechanical. The clinical accuracy pass is yours alone, and it follows the cardinal rule that governs AI progress notes: when you hand a client a document from your practice, you are attesting to its content the same way your signature attests to a note. Read every word before it leaves your office. Did simplification change the clinical claim? "Thoughts are not facts" simplified badly becomes "your thoughts are wrong," which is invalidating, or "do not trust your mind," which is contraindicated for clients with psychotic-spectrum or trauma presentations. Did the model add scope you excluded? Check the bottom of the document especially; models love to append "If you are in crisis, call..." boilerplate that may not match your actual safety plan, your state's crisis line, or what you have discussed with this client. A handout introducing crisis content you have not clinically prepared is not neutral; it is an intervention you did not choose.

Did the example fit the client in front of you? The missed-phone-call example works for most adults; it does not work for a client whose presenting trauma involves a phone call, and no model knows that. Did the homework match the treatment plan? If the plan says behavioral activation and the handout assigns thought records, the document contradicts the chart. And the hard guardrail applies here too: AI never generates risk content. No AI-drafted handout discusses a client's suicidal ideation, safety planning, or means restriction unless you wrote that content yourself and used AI only to adjust its reading level, re-verifying every claim afterward. Risk communication is clinician-authored, full stop.

One more pass clinicians skip and auditors do not: the chart. If the handout operationalizes the treatment plan, note it. "Provided plain-language psychoeducation handout on the cognitive triangle, reviewed in session, client demonstrated understanding by generating own example" ties the artifact to medical necessity in one sentence. For a payer reviewing a 90837, psychoeducation delivered and comprehension verified is exactly the kind of specific, verifiable intervention detail AI cannot supply, because only you were in the room when the client generated their own example.

Adapting for Caregivers and HCBS Settings

Now stretch the dose-form analogy one step further. In home and community-based services (HCBS) and wraparound settings, the reader is often not the client but a caregiver: the mother of a 24-year-old with schizoaffective disorder, the adult daughter coordinating services for a parent with treatment-resistant depression, the foster parent of an adolescent with complex trauma. Caregiver psychoeducation has a different active ingredient. The caregiver does not need to do the cognitive triangle; the caregiver needs to recognize what the client is practicing, support it without taking it over, and know which early-warning signs warrant a call to the team.

So the caregiver adaptation changes three things in your prompt. First, the addressee: "Write for the family caregiver of an adult client, second person addressed to the caregiver, referring to the client as 'your family member.'" Second, the content frame: what the skill looks like from the outside, how to respond when the client uses it ("you can ask, 'is that a thought or a fact?' if your family member has told you that question helps"), and what not to do (do not quiz, do not correct the client's thoughts for them, do not turn dinner into therapy). Third, the boundary content: when to contact the care team, in observable terms ("sleeping less than four hours for three nights," not "decompensation"). Reading level matters even more here, because HCBS caregiver populations span the full literacy range. Run the same Flesch-Kincaid loop on the caregiver version.

And mind the privacy line, because it is sharper here than it looks. A handout about the cognitive triangle in general is psychoeducation. A handout describing this client's specific symptoms, triggers, or treatment details, handed to a family member, is a disclosure of protected health information that for an adult client usually requires a signed release. Keep the caregiver handout generic to the condition and the skill, and keep anything client-specific inside the consented communication channel. AI makes personalizing effortless; HIPAA is the reason you sometimes deliberately do not.

Voice, Branding, and the Handouts That Should Not Exist

Two quieter points before the workflow. First, voice. A handout is a brand artifact of your practice, and AI defaults to a chirpy wellness-blog register ("You've got this!") that may clash with how you actually talk to clients. Add a voice line to your prompt: "Tone: calm, respectful, direct; no exclamation points, no cheerleading, no 'journey' language." Clients should hear your clinical voice on paper, because the handout works partly by transferring the in-session alliance to the between-session hours. Consistency of voice is part of the dose.

Second, know which handouts should not be AI-first at all. Anything functioning as informed consent (telehealth consent, AI-use disclosure, fee policies) is a legal document reviewed against your state board's requirements, not freestyled by a model. Anything addressing acute risk is clinician-authored, as covered above. Anything in a language you do not read fluently needs the bilingual verification workflow taught two lessons from now in the cultural-drift lesson, because AI translation errors in clinical material are invisible to the monolingual clinician who generated them. And anything distributed beyond your own clients, a community workshop packet or school presentation, drops the treatment-relationship context that makes psychoeducation safe, so it needs review as public health communication, including jurisdiction-appropriate crisis resources you verified yourself.

Finally, the no-PHI rule applies to handout generation exactly as it applies everywhere else. You do not need the client's name, initials, or identifying narrative to generate a handout, so none of it goes into the prompt, even in a BAA-covered tool. "An adult client in treatment for depression" is all the model needs. The moment you find yourself typing the client's actual story into a prompt to make the example fit better, stop, abstract the pattern ("an example involving a coworker conflict"), and let the model invent the specifics. Personalize by hand afterward if clinically warranted.

The Applied Problem: Your Reading-Level Handout Workflow

Your artifact is a one-page, reusable document titled the Reading-Level Handout Workflow, used every time you generate client-facing psychoeducation and shown to a supervisor or compliance officer as evidence of a defensible process. Build it in four sections. Section one is your prompt template with five fill-in slots: TOPIC, CLIENT POPULATION (never an identifiable client), READING LEVEL with the Flesch-Kincaid target number, STRUCTURE (main point first, one concrete example, one practice step, bring-questions-to-session closer), and EXCLUSIONS (no medication content, no diagnosis labels, no crisis content, no additions beyond the topic, plus your voice line). Paste in the cognitive-triangle prompt from this lesson as your worked starting point.

Section two is the measurement loop, three numbered steps: (1) paste output into a Flesch-Kincaid checker; (2) if the score exceeds target by more than half a grade, send the score back with the split-sentences-and-swap-syllables instruction; (3) re-measure; do not proceed until the number is right. Section three is the clinical accuracy checklist, six yes/no items you initial: claim survived simplification unchanged; no added scope at top or bottom; example safe for this client; homework matches the treatment plan; no risk content unless clinician-authored; no PHI entered the prompt. Section four is delivery: review the handout with the client in session rather than at the door, and write the one-line chart note tying it to the plan.

Test the workflow now, before you need it. Pick a concept you teach weekly, sleep hygiene, paced breathing, the window of tolerance, and run it through all four sections twice, at 6th grade and at 10th. The first run takes twenty minutes; the fifth takes six. Done looks like this: two versions of one concept, both scored and converged on target, both passed through the six-item checklist with your initials, a caregiver variant generated if you work in HCBS or family-involved settings, and the workflow document saved where you can reach it during a documentation block.

Key Takeaways

• A psychoeducation handout is a dosed intervention: the clinical concept is the active ingredient, the reading level is the dose form. Roughly one in five American adults reads at or below a 5th-grade level, so a handout pitched above the client's reading level cannot be absorbed.
• AI collapses the cost of producing the same concept at multiple reading levels, 6th-grade and 10th-grade versions of the identical cognitive-triangle content, but introduces three recurring defects: reading-level drift, clinical drift during simplification, and scope creep such as unrequested crisis or medication content.
• Never trust the model's self-reported reading level. Verify with an independent Flesch-Kincaid check, and use the fix loop: report the actual score back, instruct the model to split sentences over 14 words and swap multisyllabic words, then re-measure until you converge on target.
• CMS plain-language standards check what formulas cannot: main point first, active voice, content organized around what the reader does. A handout can hit grade 6.0 and still fail plain-language review. Medicaid managed-care materials often carry an explicit 6th-grade target.
• The clinical accuracy pass is non-delegable. Confirm the claim survived simplification ("thoughts are not facts" must not become "do not trust your mind"), the example is safe for this client, the homework matches the treatment plan, and no risk content appears unless you authored it yourself. You attest to every client-facing document the way you attest to a signed note.
• Caregiver versions for HCBS settings change the addressee, the content frame (recognize and support the skill, do not run it), and add observable when-to-call-the-team criteria, while staying generic enough that no client-specific PHI is disclosed without authorization.
• No PHI ever enters a handout prompt, even in a BAA-covered tool: describe the population, not the person. Personalize by hand afterward if warranted, and chart the delivery with one sentence linking the handout to the treatment plan and the client's demonstrated comprehension.