Hire the smartest clinical assistant in the world.
Meet Clay.
Clay is the world's smartest clinical assistant. He has access to the peer-reviewed research on your modality, the emerging interventions, the leading science of how people change, and he hands you a grounded answer in minutes.
He reads a de-identified version of your new intake through your modality and hands you the structured read before your next session. He drafts the 9pm follow-up in your voice. He scores the assessments, holds twelve sessions in his head at once, and shows you the pattern you could feel but could not name. Ask him, and he will tell you, directly, how you did in the room.
And he cannot break your oath. Clay never meets your clients, so he cannot compromise your ethics: every identifier is stripped by his machine, the Claudette, on your own local computer, before he reads a word. Nothing identified ever leaves for a cloud. Not a policy he follows, an architecture he cannot see past. What he gives back is more than your evenings. It is the fuel to be fully human with the humans in front of you. Will you stay in the chair?
Things you ask Clay.
Real requests, run on de-identified, code-only material, answered in the modality you taught him. Say it in plain clinical English; he does the rest.
"Run the total analysis on PT-0114. Put her baseline anxiety from intake next to what she self-reported last session, in a table, with the turning points between."
"Give me a one-page summary of each partner's attachment style, with the transcript moments that show it. Something I can print for my own file."
"Map this couple's pursue-withdraw cycle from tonight's session: the triggers, the secondary emotions, the attachment fear under each move."
"Lay out PT-0207's automatic thoughts from the last four sessions, the distortions they map to, and which restructuring work we have not tried yet."
"List the parts that showed up in session twelve, what each one seems to be protecting, and the places I talked to a part as if it were the whole person."
"Read the whole case and name the unconscious patterns: what keeps repeating in how he leaves things. Jobs, sessions, sentences."
"Summarize skills use across her diary cards: which distress-tolerance skills she actually reaches for, which she avoids, and what that says about the next module."
"Track the body notes across sessions: where he goes still, where the breath changes, and which topic comes up right before it, every time."
"Sort what she did this month into values moves and avoidance moves, in her own words, and show me where the avoidance is winning."
"Tell me where I led the witness this week. Quote me."
"Draft the between-session email that holds what she said at the door, in my voice, without making it bigger than she made it."
"What is the current thinking on exposure pacing for panic with a comorbid breathing condition? Grounded answer, with the sources I should check."
If your modality is not on this list, that is the point. He thinks in whichever one you teach him, and teaching him is not a research project you run alone: the guide hands you the exact prompt, you copy it, you paste it, and you receive an up-to-date reference for your modality. Every request above runs on material the Claudette cleared first: codes in, codes out, judgment yours.
You are afraid of the wrong thing.
You are locked out of the best tools by the best thing about you. The oath is not the obstacle; it is the reason this exists.
Here is the double bind nobody names. You are told AI is coming for your profession, though it cannot sit in the chair and the studies say exactly that if you read them instead of the headlines. And in the same breath you are shown tools you cannot touch, because you made a promise. Client information does not go where it is not protected. Your clients depend on that promise. It is not a compliance checkbox; it is a load-bearing part of why therapy works at all.
So the most careful clinicians, the exact people who should be leading here, sit it out. Not from ignorance. From integrity. And every month they sit out, the gap widens between what their work could be and what their evenings look like.
The way through was never to loosen the promise. It is to make the material genuinely not-about-anyone before any tool sees it. De-identified to the Safe Harbor standard, a session stops being a client's file and becomes a case study, one you can run through your own modality, your own frameworks, your own analytical lens, with the smartest assistance available on the planet at your disposal.
That does not just make you faster. It makes you a well-resourced clinician who does not miss very much: the pattern across eight sessions, the thing they said three times, the question you never came back to. The promise stays whole. The work gets sharper. Both, not either.
AI is the multiplier, not the replacer.
A licensed clinician, not a prompt guru.
You on camera, fixed setup, no B-roll. Say who you are, that you are licensed and in practice, and that you built this on your own clinical work before asking anyone else to trust it. Anchor it in the 9:40pm moment: four notes left, the tab open, the finger hovering.
Do not tell a transformation story and do not sell in this one. Land on the line the whole page hangs off: I am going to show you the line between the hype and the fear, drawn precisely.
You cannot paste a client's intake into a chat window.
Correct. And it stays correct.
Any tool that handles protected health information on your behalf is a business associate, and it needs a signed agreement in place before a single identifier moves. Consumer tiers do not offer one. No badge on a homepage changes that, and no vendor's marketing copy will hold up in front of a board.
So here is the sentence the rest of this is built on. There is no such thing as a HIPAA-compliant AI. Compliance is not a feature you can buy. It is a property of how protected information is handled, which means the same tool is either defensible or career-ending depending entirely on what you did in the half second before you pressed enter.
The protection was never in the software. It was always upstream, in you, in the discipline of removing a person's identity from the material before the machine ever sees it. Once material is de-identified to the Safe Harbor standard it is no longer protected health information, and it sits outside the Privacy Rule. That is not a loophole and it is not a workaround. It is the rule, used the way it was written.
The problem with leaving that to your own care is that human care fails at 11pm on a Thursday. So you do not leave it to your care. You run a program that removes the identifiers, checks its own work afterward, and refuses to release a file that still carries one.
And there is a second lane, for the one step where identified material has to touch a service at all. Some companies will sign a Business Associate Agreement with a clinician, which makes them legally accountable for protecting what you send. Transcription is the step where this matters: this build uses a transcription vendor precisely because they sign a BAA, and the guide names what any vendor must offer before audio goes anywhere: a signed BAA, speaker separation, a written opt-out of training on your data. Two lanes, one rule. PHI only ever touches a party bound to protect it, and everything downstream of the boundary is not PHI at all.
The whole workflow, on one map.
However your material arrives, it drops into the same process. Follow any line from top to bottom: nothing crosses the boundary carrying a name, and nothing reaches a client without passing through you.
The same map, step by step.
Step four is a program, not a promise. Step seven is the one that is not optional.
-
01Identified · your machine
The material arrives.
A recorded session, an intake, your own notes, an assessment. Whatever you already have. It lands in one folder and it does not move until you move it.
-
02Identified · your machine
Audio becomes text.
Transcription with speaker separation, so you can tell your voice from theirs. There is a fully local route where the audio never leaves your computer, and there is a faster route with the requirements a vendor has to meet before you send it anything.
-
03Identified · your machine
You say who it is about.
The tool never guesses. You give it a case code or the names, or it refuses to run. That refusal is a feature: a de-identifier that guesses who a document concerns is a de-identifier that will eventually guess wrong.
-
04The boundary · offline
De-identification, by program.
Two passes. First the things that have a shape: emails, phone numbers, addresses, dates, record numbers, ZIPs, ages over 89. Then the names, which have no shape and get decided on evidence rather than vibes.
Then Safe Harbor's eighteenth category, the one most home-made scripts miss entirely: any other unique identifying characteristic. In a small metro that includes a medical subspecialty, an unusual hobby, a relative's profession, and a dollar figure.
Ambiguity always resolves toward redaction. No model, no network, no account. You can read every line of it, and you should.
-
05The verification gate
The file is checked again, by different code.
Every output is re-scanned before it is allowed through, by detectors written independently of the ones that did the stripping. If one implementation has a blind spot, importing the same pattern to check its work would miss it twice.
Anything that still carries an identifier goes to quarantine with a note saying what was found and where. A file that leaks is never released quietly.
-
06De-identified · safe to work with
Now you think out loud.
The material that reaches the model is not about anyone. So you can ask for the formulation in your own modality, the three hypotheses with what would confirm and what would disconfirm each one, the thing conspicuously missing, the shape of a first session, a draft in your voice.
Then you argue with it. The value is never in the first answer. It is in the third.
-
07You, and only you
You read it, and you decide.
Nothing in this system sends anything. Nothing files to a client record. Nothing signs your name. Output comes back to you, you read it, you keep what is true, and you re-identify and send it yourself, on your side of the wall.
That is a deliberate limit and you should not engineer around it. The moment something goes out without passing through you, you own a document you did not read.
This is the Claudette's work: material goes in identified on the left, comes out a case study on the right, and the machine never touches the network doing it. Every person, place and employer in this example is invented. The clinical meaning survives, which is the actual difficulty: a de-identifier that destroys the meaning is just a shredder. Note the sister's specialty went too. In a town with one of those, a name is not what identifies her.
If you can read it, she can clear it.
Anything that lives as text can become not-about-anyone, and from there it is a case study you can think with. And when she cannot clear something, she does not shrug: she quarantines it and tells you why. That refusal is the feature.
Two sides, one wall, one tool that crosses it.
You stop relying on your own vigilance and start relying on where a file physically lives.
Your side.
Under your control, treated exactly the way you already treat chart notes.
Claudette/ her typed trays: audio, transcripts, intakes, assessments, notes, email Claudette/vault/ originals plus the crosswalk, hers alone Claudette/quarantine/ files that failed the check, reason printed _registry.json the code-to-person map. This file is PHI.The working side.
Codes only. No names, no dates, no places. Safe to reason over.
Clay/cleared/ de-identified source material, the only folder he reads Clay/analysis/ what he brings back, filed by case code Clay/drafts/ client-facing drafts, unsentWatch an identifier disappear.
Screen recording, no face, voiceover only. A synthetic intake on screen carrying a name, a town, an employer, a date, a specialty and an email. Run the command. Put the cleared file beside it so the codes and the surviving clinical meaning are both visible in one frame.
Then show the gate doing its job: a file that still carries something goes to quarantine with the reason printed. Say it once, out loud, while the network indicator sits idle: nothing left this machine.
Invented material only, never a real client file, and never a mock-up of anybody's product interface.
Clay is the face. The Claudette is his machine. Claude is the brain.
One assistant, one instrument: the Claudette guards the boundary, and Clay does the thinking on the other side of it. Here is what he is made of.
Claude, under the hood
Clay's reasoning runs on Claude, and the guide walks you through that setup step by step, because that is where the depth for long clinical documents currently lives. And because the Claudette sits upstream of the brain, you are never trapped: what she releases is safe in front of any model you point him at, so if the landscape shifts, the brain changes and the boundary never moves. You bought the assistant, not a subscription to a vendor's roadmap.
He thinks in your modality
This is not a generic AI bolted onto therapy. You teach him your way of working: a CBT therapist builds a CBT reference and gets formulations that cite the model; an EFT therapist gets cycle mapping; an IFS therapist gets parts language. You build your modality document, your voice document, and your own analysis workflows from elicitation prompts included in the guide, each one written out in full so none of this asks you to invent anything, and everything he produces reasons in your lens and sounds like you. You control what gets analyzed, from what perspective, by what method.
He learns how you correct him
The part that compounds. When he gets something wrong, you do not fix it twice. Say the correction once: never use that word, never frame it that way, always check scoring bands against my reference, and it goes into a standing rules file he reads before every piece of work, forever.
Month one he knows your modality. Month six he knows your judgment: the corrections you would give a bright intern, permanently kept. Nobody else's assistant gets smarter about you over time.
Analyze. Draft. And the one nobody advertises.
Your thinking partner for compliant efficiency: it multiplies your competence and leaves you the fuel to curate human-to-human experiences with your clients. None of it replaces you; every piece of it requires you to read the output and decide. What it replaces is the cold start.
Analyze
A new intake arrives and you have eleven minutes before your next session. Instead of skimming it, you sit down to a structured read: what they are presenting with, what the pattern suggests, what is conspicuously missing, and four questions worth asking in the first twenty. Grounded in your modality, because you taught it yours.
Draft
It is 9pm Thursday and you owe a between-session email to someone you saw on Tuesday, and you cannot find the sentence. Give it the de-identified notes, get a draft in your voice holding the thing you actually wanted to hold. Rewrite two lines and send it. Fifteen minutes instead of fifty, and the fifteen were the good part.
Evaluate you
When was the last time anyone watched you work? Not the hard session, the one you brought to consultation. An ordinary Tuesday. For most of us the honest answer is graduate school.
Point it at a de-identified transcript of your own session and ask how you did. Where you interrupted. Where you reached for technique instead of staying with the affect. What they said three times that you never came back to. There is always something.
Why the insight is not the work.
You on camera. The model hands back a clean formulation in eleven seconds, and it reads well, and a clinician who mistakes that for the work gets worse rather than better. A hypothesis is not a finding. Fluent is not correct.
Show the posture instead: you push back, you say why you disagree, you ask what it is missing, you ask it to make the strongest case against your own formulation. Land on the rule that governs everything on this page. It produces hypotheses. You decide what is true.
The clinical assistant you could never afford.
Open cold on the question, no runway: when was the last time anyone watched you work? Not a hard session. An ordinary Tuesday. Wait for it to land before you answer it.
Then the evaluate-you workflow running on a de-identified transcript, and be specific about what came back: where you interrupted, where you went to technique, the thing raised three times you never returned to. Say honestly that you did not like all of it. That honesty is the whole reason this one converts.
143 pages. Every step written out. Scroll a few of them.
Real pages from the Guide, prompts blurred until it is yours. In your copy, every prompt and command carries a one-click Copy button: copy, paste into your AI, go.
Workflow A — Intake analysis
Open your Project. Paste the de-identified intake and this:
Intake for [CASE CODE]. De-identified.
Analyze it using the frameworks in my knowledge folder. Produce:
You still decide what is true. You just did not start from a blank page.
Workflow E — Evaluate me
The one nobody else offers. Give it a de-identified transcript of your own session:
This is a de-identified transcript of a session I ran. Evaluate me, not the client.
Module 04 — Configuring the Claudette
Your protected phrases, your local gazetteer, your case codes. The twenty minutes that make everything else safe:
Then the test suite: a hundred-plus checks on your own machine, under a second, after every change.
Appendix — the complete source
Every line of the Claudette, printed in full, so you can read what you are trusting before you trust it.
The blur comes off at checkout. The Toolkit tier adds all of this as working files, so there is nothing to type at all.
Who this is for, and who it will hurt.
Built for
- A licensed clinician in private or small-group practice who does their own notes, their own intakes, and their own follow-up.
- Any modality. You teach it your lens in the third module and everything downstream thinks in that lens.
- Any comfort level with technology. Module 02 assumes you have never opened a terminal and explains what one is before asking you to use it.
- Anyone who has been curious about this for a year and has not touched it, because nobody would tell them where the line was.
Not for
- Anyone looking for AI to write notes they will sign without reading. You will get yourself in trouble and I would rather not have helped.
- Anyone who wants a tool that puts something in front of a client. That is a different product, in a year when three states have already legislated against parts of it.
- Anyone hoping to skip the boring twenty minutes where you list the towns and hospitals near you. That step is what makes the rest of it safe.
- Anyone who wants a certificate that says they are compliant. No such certificate exists, and the ones being sold are worth what they cost.
Empty folder to working system.
Screen recording with voiceover, sped up where nothing is happening. Start on an empty folder. Build the structure, write the project instructions in a modality, configure the tool with a gazetteer, then run one invented intake all the way through to a formulation on screen.
Keep a clock visible. One focused Saturday, compressed into three minutes. This is the video that sells the toolkit tier, because it shows the exact tedium the toolkit removes.
A clinician built this. That is the whole point.
Christian, plain setup,
same framing as the videos
I am Christian J. Charette, a licensed marriage and family therapist in active practice. I built this system for my own caseload first: my own intakes, my own late notes, my own pattern-hunting across sessions, and I ran it there before asking any other clinician to trust it.
I built it under two commitments that do not bend. The first is the oath. Our clients hand us the most identifying material a person has, and they do it because they believe it stays protected. Nothing in this system asks you to spend that trust; the entire architecture exists so you never have to. De-identification is not a feature here. It is the foundation everything else stands on, held by a program rather than a promise, because programs do not get tired at 11pm.
The second is the craft. I did not want AI to do my thinking; I wanted to become a better clinician with the best-resourced thinking of my career. The smartest assistance on the planet, reading my de-identified cases through my modality, in my frameworks, drafting in my voice, showing me what I missed and telling me directly how I did. That is priceless, and it was locked behind a wall for every clinician who takes the oath seriously.
This is the door through the wall. Clinician to clinician, with the receipts to check.
One purchase. It runs on your computer.
No subscription, no seat, no portal. It keeps working if you never hear from me again.
Guide plus Toolkit
What’s inside
- The Claudette as a working folder: configured by template, documented, with a hundred-plus-check test suite that runs in under a second on your machine.
- The workspace scaffold, unzipped into place instead of typed.
- The elicitation prompts as files: your modality reference, your voice document, your local gazetteer.
- All seven workflows as installable skills: intake analysis, session analysis, assessment review, the client-facing draft, the longitudinal look, evaluate me, and scoring for your own instrument.
- The standing-rules file, shipped and wired into every skill: correct him once, “never do that again,” and he keeps it, in every future session.
- Bonus: twenty modality packs, one-copy installs: CBT, DBT, ACT, EFT, IFS, EMDR-informed, psychodynamic, and more. Pick yours from the list, drop it in, and Clay reasons in your lens on day one, no research weekend required. Plus the pack builder: if yours is not among the twenty, one copy-paste prompt researches your modality and writes its pack for you.
- Bonus: Build Your Own Assessment. Module 07 turns the questions you already ask into a scored instrument with your name on the scoring document, plus the skill that scores it and feeds it into the longitudinal look.
The Guide
What’s inside
- Nine modules, 143 pages, in order: one focused Saturday to a working system, plus the standing-rules, assessment, and mobile-drop bonuses.
- The compliance base, written in language a licensing board would recognize.
- Every prompt one-click copyable in the interactive edition: copy, paste into your AI, go.
- The Claudette’s complete source, printed in the appendix, so you can read every line before you trust it.
- Written for a clinician who has never opened a terminal window.
Clinician to Clinician
What’s inside
- Two one-on-one working sessions with Christian over Zoom, on your actual machine, your actual practice.
- Your workspace stood up together: folders, the Claudette configured with your local gazetteer, the boundary tested live.
- Your modality and your voice document elicited and installed, so it thinks in your lens from day one.
- Your first real workflow run end to end on an invented case, with you at the keyboard.
- Licensed clinician to licensed clinician. No screen-share support tech. Someone who does this work daily.
The difference between the first two tiers is typing: the Guide prints everything so you can build by hand and understand it completely; the Toolkit hands you the same system as folders that already run, so the single most common failure, a non-technical buyer mistyping one character in a file that handles PHI, cannot happen to you. The third tier is for the clinician who wants it stood up right, once, with someone who has done it on a real practice.
What this will not do.
Read this part twice. It is the part that filters out the buyer who gets hurt.
- It will not make clinical decisions. It produces hypotheses. You decide what is true, and when it is confidently wrong in fluent, convincing prose, you are the only safeguard in the room.
- It will not write notes you do not read. If you are looking for documentation you can sign unread, close this page.
- It will not make you compliant. It gives you strong tools and a defensible architecture. Your agreements, your risk analysis, and your Safe Harbor determination stay yours.
- It will not replace supervision or consultation. It is a thinking partner, not a licensed one, and the difference is not cosmetic.
- It will not keep you current on its own. Three states legislated on AI in therapy inside a year and forty-three more introduced bills. The guide is dated, and staying current is your standing job.
The ones that come up first.
Do I need to be technical?
No. Module 02 assumes you have never opened a terminal window and explains what one is before asking you to use it. Every command is written out in full to copy and paste, and where something commonly goes wrong the fix is printed right there rather than buried in an appendix. If you get stuck, the instructions failed, not you.
Does this work with my modality?
Yes, and that is the design. Nothing in it is bound to a school of therapy. In the third module you teach the system your own frameworks, your own reference material, and your own clinical language, and everything downstream reasons in that lens rather than a generic textbook one. CBT, DBT, EMDR, IFS, EFT, psychodynamic, couples, addiction. You supply the lens.
Does any client information get uploaded?
No. The de-identification tool has no network code in it at all. No account, no API key, no dependencies to install. It reads a file on your computer and writes a de-identified copy on your computer. Only the de-identified copy is ever used with an AI service, and the crosswalk that could reverse it never leaves your machine.
What about recording and transcription?
Covered as an optional module, with both routes. There is a fully local route where the audio never leaves your computer, which is slower to set up and answers the question completely. There is a faster cloud route, and the module gives you the requirements a vendor has to meet before you send it anything: a signed business associate agreement, speaker separation, and a training opt-out. Verify those directly with the vendor. Never take them from a webpage, including this one.
Which AI does it use?
The guide is built around Claude, because that is where the long-document reasoning and the depth currently sit, and the walkthroughs are written against it step by step. The method itself is model-agnostic. Once material is genuinely de-identified, the pipeline in front of it is the part that protects you, not the brand of model behind it.
Is this legal advice?
No. It is an educational product built by a clinician, not by an attorney, and it does not certify anything. It teaches Safe Harbor by the federal standard and gives you a defensible architecture. Your business associate agreements, your risk analysis, and your own de-identification determination remain yours to make and yours to own.
Is there a refund?
Read the section above this one first. It is written to talk the wrong buyer out of purchasing, on purpose, because the wrong buyer here is a clinician who gets hurt. If you read it and it still is not what you thought, write to me and I will refund it.
Leave an address. Stay close.
One letter when something real ships: a new module, a new pack, a change worth knowing about. No sequence, no drip, and unsubscribing is one click.
Nothing clinical ever goes to email. Yours or anyone's.
Addition, not subtraction.
The finger hovering over the tab at 9:40 was never the problem.
It was the right instinct, reaching the wrong way. Here is the right way to reach.