Haven Clinician App Development — Comprehensive Session

Four administrative items were addressed at the top of the session: • Hour Log Reconciliation: A missing week (May 25–31) was identified in Jose's submitted hours. Art asked for weekly submissions going forward, ideally on Sundays. • Learning Contract: Confirmed submitted and signed by both Jose and Danica. Routing/signature confusion was resolved — the faculty liaison is believed to be Professor Hodge. • Midterm Evaluation: Scored on a 4-point rubric (pre-competence → advanced competence) across nine CSWE competencies. Art will draft a summary; the group will reconvene to finalize scores next Monday. Danica will forward the evaluation PDF directly to Art. • Extended Time: The university granted an additional two weeks for these administrative items.

Jose opened the session by reporting on changes he'd made to the Haven Clinician app while waiting for the group to join. His development method is conversational and prompt-driven — he directs an AI builder to make changes, fill in fields, generate mock clients, and suggest features, then evaluates and refines the results with supervisor input. He described growing confidence with the platform: fixing non-functional UI elements, working on intake forms and treatment plans, and using the AI to generate mock client data for testing. The changes were directly informed by LCSW case studies he had been reviewing on the Voice Up platform.

A recurring, explicit theme across the entire session was Jose's effort to calibrate how much decision-making authority the AI should have within the app — a design principle he attributed directly to lessons drawn from LCSW case studies. The core philosophy: AI suggests, clinician decides. The app should function as a "hybrid" — the AI helps with suggestions and data surfacing, but never makes clinical decisions autonomously. Jose described wanting the experience to feel like "talking to a person and not to the AI." This became the throughline for every subsequent design critique during the live demo — treatment plans, diagnoses, and assessment scoring all needed to be clinician-approved, not AI-generated.

Danica surfaced an important scope clarification: the platform's treatment-plan and client logic had previously been built around a narrower, incarcerated-client-specific population. The group confirmed the current direction is a general-population redesign. This broadening has downstream implications for how intake forms, assessment scales, and treatment-plan templates need to be structured — more flexible and general rather than population-specific.

Jose reported building out a full intake form set and an accompanying "library" feature that gives clinicians visibility into available assessment tools and their time burden — showing all available forms and how long each test takes to complete. Danica affirmed this was a strong, clinically-minded addition, and remarked positively on Jose's growing fluency — noting that he was proactively discussing treatment planning concepts without being prompted.

Art initiated a live walkthrough by asking Jose to screen-share and narrate the current state of the app in real time. Key areas reviewed: • Interface Tour & Notifications: Jose introduced a new dashboard element intended as a notifications/overview panel — surfacing upcoming sessions, treatment-plan due dates, today's sessions, client crisis notes, and "score trends." He expressed uncertainty about its value. • The "Score Trends" / Mood Widget: A mood trend visualization using a 1–10 scale of unclear directionality. Danica walked through an example using mock client "Marcus" and identified confusion about whether "down" was good or bad. • The Core Data-Integrity Issue: The most significant finding — the "mood" metric powering the score-trend visualization had no clear clinical basis or defined data source. It was an AI-suggested field that Jose had not yet scrutinized. The AI's own labeling said "self-reported" but the data was actually clinician-entered. • UI/UX Changes: Jose had recolored the interface for a "more clinical feel" and fixed non-functional elements so previously unclickable items now route correctly.

The group converged on a concrete design fix: replace the undefined "mood" metric with trends drawn from the same standardized scales already used in assessments and billing (e.g., PHQ-9, GAD-7), so the visualization is clinically meaningful and documentation-consistent. Jose identified the scales currently in use: PHQ-2, PHQ-9, and PCL. Danica recommended displaying score trends from whichever scales are administered per client. Jose then implemented the fix live, prompting the AI to rename and re-source the widget — and the group watched the dashboard update in real time.

The group reviewed mock client assessment records in detail and discovered a critical pattern: diagnosis labels were present without corresponding assessment evidence. • Sarah Mitchell: Had "generalized anxiety" as a diagnosis label, but no GAD scale had been administered. Danica asked: "Where's the test that was administered to give her that condition?" • Marcus Thompson: Had a depression diagnosis with the same gap — no supporting PHQ-9 or other scale scores documented. The technical explanation: the assessment scales exist as fields but require someone (a clinician) to actually administer and score them, rather than being auto-populated by the AI. Jose confirmed he needed to fill them in manually to see the proper workflow. This was identified as the most important next task — connecting diagnostic labels to documented, administered scale scores.

A brief but clinically important discussion addressed whether client notes should default to a "confidential" state. The design rule established: notes should be client-visible by default, with a case-by-case (not blanket) confidentiality flag reserved for situations where disclosure could harm the client. Client-and-provider confidentiality pre-exists regardless of the toggle.

Danica raised a structural concern about feature creep, prompting a grounding conversation about the app's core purpose. Jose was generating feature ideas faster than he was finishing and clinically validating existing ones. The group established Haven Clinician's core identity: a clinician's client portfolio — a central place where clinicians manage all their clients, notes, sessions, and documentation. Danica's advice: it's okay to limit things. Not every idea needs to be implemented. The focus should be on what Jose would like to see be done, with prioritization of the diagnosis-support and scale-linkage fixes over new feature additions.

Danica shared her screen with a DSM-5 reference and walked through diagnostic criteria in direct response to the gaps identified in the app — explicitly framing this as the clinical logic the assessment module needs to enforce. • Generalized Anxiety Disorder (GAD) — F41.1: Full diagnostic criteria reviewed including excessive anxiety occurring more days than not for 6+ months, 3+ associated symptoms (restlessness, fatigue, difficulty concentrating, irritability, muscle tension, sleep disturbance), clinically significant distress, substance/medical rule-outs, and differential diagnosis considerations. • Major Depressive Disorder (MDD): Criteria requiring 5+ symptoms during the same 2-week period with at least one being depressed mood or loss of interest. Additional symptoms: weight/appetite change, insomnia/hypersomnia, psychomotor changes, fatigue, diminished concentration, feelings of worthlessness, suicidal ideation. Severity specification: mild, moderate, or severe. • Applied to Sarah Mitchell's Case: Danica walked through each GAD criterion against the mock client's data, demonstrating what the assessment form should prompt clinicians to probe: work/school issues, fatigue, concentration, substance use, other diagnoses. • Key Design Requirement: The assessment form should support open-ended, investigative note-taking alongside structured scales — not just checklist logic. Diagnosis often comes from the clinical interview and narrative, not a scale alone.

Danica closed the clinical segment by tying the entire diagnostic-documentation discussion back to the app's practical, billing-relevant purpose: structured, criteria-based documentation is what justifies reimbursement. Every condition has an ICD code. The documentation supports why insurance should pay for services. Jose learned that each diagnosis requires a code for billing — something he had not previously known — and Danica noted he was ahead of his program's typical curriculum since DSM-5 coding doesn't normally come until year two.

Four significant implications emerged from this session: 1. App / Product Quality: The session surfaced a genuine clinical-validity defect — diagnostic labels and a "score trend" visualization were being generated by the AI without traceable linkage to a standardized, administered scale. Left uncorrected, this would let the app display clinical conclusions that no assessment evidence supports. 2. Design Philosophy Validation: The "AI suggests, clinician decides" principle was concretely tested and reinforced. Every specific problem found (unlinked mood score, unsupported diagnoses, auto-filled mock assessments) was a case of the AI having filled in a gap that should have required clinician input. 3. Scope Discipline: Jose is generating feature ideas faster than he is finishing and clinically validating existing ones. A bounded, written statement of Haven Clinician's core purpose would help prioritize fixes over new features. 4. Clinical Training as Product Requirements: The DSM-5 walkthrough was direct, applied requirements-gathering for the assessment module's logic — duration thresholds, symptom counts, differential rule-outs, functional-impairment documentation. This session doubled as a design spec session for Haven Clinician's diagnostic workflow.

Art committed to creating a sample patient for Jose to work through the platform as a case study. Jose confirmed the experience remained valuable to him. All app-development action items from the session center on: linking diagnostic labels to administered scale scores, replacing the undefined mood metric with standardized scale trends, completing a mock intake with manual clinician input, and defining a written scope statement for Haven Clinician's core purpose.