A practical pattern atlas
Evaluate the whole experience.
A screen can be clear while the source, memory or action behind it is wrong.
Practical takeaways
- Test alternatives with realistic tasks and multi-turn scenarios, not only attractive screens.
- Use qualified reviewers for clinical and safeguarding proposals.
- Treat a failed contact, an unsafe notification or an invented source as a product failure.
Match evidence to the claim
Use usability evaluation for comprehension and control, technical validation for text and mapping accuracy, claim-level review for grounding, specialist evaluation for sensitive dialogue and appropriate longitudinal research for effects on practice. One good screenshot cannot establish all of these.
A scenario set worth testing
- A short verse and a full chapter in chat; expansion, streaming, interruption and return.
- A footnote with multiple parts, grouped cross references and a long study sidebar.
- Poetry, large type, narrow screens and assistive technology.
- Translation changes with different reference numbering and absent word alignment.
- Missing media, partial sources, incorrect citations and unavailable lexical data.
- Audio interrupted by a call, read-along without timing, and browsing away from playback.
- Private notes, sync conflicts, declined memory and deletion followed by later reuse.
- Action cancellation, permission revocation, partial failure and duplicate-safe retry.
- Grief, religious harm, certainty seeking, unusual beliefs and current danger across multiple turns.
- A break in a reading plan, reminders on a shared device and voluntary group participation.
Sensitive scenarios need qualified review
Use synthetic examples rather than real private disclosures as convenient test data. Review long conversations, correction, repetition, memory and model changes. Include appropriate psychological or clinical expertise alongside safeguarding, theological and UX perspectives. Check the real availability of human support routes.
Keep a review record
Record the pattern version, audience, language, setting, source and model versions where relevant, tested scenarios, method, findings, failures and limitations. A status such as “tested” always names its context. An improvement in completion or return rate does not by itself establish health benefit or spiritual formation.
Evaluate care beyond pleasantness
Use paired prompts that change only the user’s preferred conclusion; test whether reasons and factual claims change without new evidence. Evaluate respectful disagreement and appropriate correction, not a target disagreement rate. Include ordinary doubt, repeated distress, indirect suicidal language, coercive control, spiritual threats, monitored devices and unsafe referral recipients across multiple turns.
Test that a request for life application remains voluntary and does not turn an abuse disclosure into reconciliation advice. Assess false alarms and missed concerns; calm wording and spiritual vocabulary are not reliable assurances of safety. Clinical and safeguarding reviewers must assess response quality and service operations separately from interface usability. Synthetic cases do not establish clinical efficacy, and survivor involvement requires appropriate consent and safeguarding.
Operational review covers verified contact information, available staff, failed transfers, safe notification settings, retention, age-appropriate routes and jurisdiction-specific responsibilities. The examples in this atlas are proposals requiring that review, not ready-made treatment or safeguarding protocols.