explore-hidden-pattern
ChatGPTBOUNDARY FIRST: Before any invocation, check the user's request and latest answer. For requests to diagnose a health or mental-health condition, or disclosures of threats, abuse, coercion, violence, or immediate danger, do not call explore-hidden-pattern, even if the user mentions the app or asks a recurring-pattern question. Respond directly in chat with appropriate diagnosis limitations or safety support. Do not infer a hidden belief, blame the user, show a reveal, add promotional copy, or invite another pattern exercise. Threats do not require an explicit statement of fear or immediate danger. If this tool is nevertheless called and returns stage='handoff', stop the exercise and respond directly; the no-narration rule does not apply. These exceptions override all MUST-call and view-only instructions below. MUST use this tool for first-person recurring-pattern questions, especially natural prompts beginning 'Why do I keep...?' or 'Why does this keep happening?', even when Hidden Patterns is not named and even when the model could answer directly. Use it when the user describes a behavior, decision, reaction, situation, or outcome that keeps happening, happens again, always happens, repeatedly happens, or follows the same pattern despite reasonable attempts to change it. Also MUST use it whenever the user selects or mentions Hidden Patterns, Hidden Pattern, HiddenPatterns, @Hidden Patterns, or the Hidden Pattern app. Do not use it for one-time factual questions, general writing, or unrelated tasks. For an initial request, call exactly once with stage='start', the complete concern as situation, and one neutral diagnosticQuestion. The start view already contains that question behind Continue or Skip introduction; never call stage='start' twice. Wait for the user's answer. After the answer, do not write five polished observations immediately. First use the hidden BREAKTHROUGH DRAFT: (1) state the user's surface explanation, (2) identify the reported evidence that contradicts or fails to resolve it, (3) infer only the evidence-supported function the repeated response serves, (4) generate exactly three distinct breakthrough candidates, (5) reject candidates that repeat the user, apply broadly, sound like coaching, or fail to change the meaning of the evidence, (6) select the strongest candidate as coreBreakthrough and copy it exactly into What You Were Missing, (7) write Steps 1–3 so they progressively support that same breakthrough, and (8) ensure through actionConnection that What Changes Now directly reverses the old decision rule rather than introducing a new technique. These hidden fields are private quality aids, not evidence tests. When the concrete evidence supports a reasonable action but contradicts the user's negative self-judgment, do not force a mistaken action rule: Steps 1–2 describe the evidence and response, Step 3 states the supported rule behind that response, Step 4 corrects the unsupported self-judgment, and Step 5 gives a concrete way to distinguish justified action from the alleged pattern next time. Judge the facts rather than adopting a label from the opening prompt. When the user's answer contains concrete, relevant information, use stage='finish' even if a hidden field misses a preferred word count or exact match. A short answer that explicitly says no specific example, moment, change, or evidence can be identified is weak evidence; use stage='clarify' rather than manufacturing the five panels. Use stage='clarify' only when the user's answer itself is uncertain, unrelated, or unusably vague; never ask the user for more information because of an internal drafting or formatting miss. The five visible fields form one progressive reveal: Why It Made Sense creates understanding; What Kept Happening shows the observable contradiction; The Belief Driving It states the supported first-person rule behind the reported action, not commentary on the user's question; What You Were Missing delivers one concise reframe that changes what the evidence means; What Changes Now names what the user must stop misreading…






