Purpose: Five things to ship this week. Five things to ship this month. A list of anti-patterns to audit and remove. This is the implementation companion to the PhD memo.
---
What: Drop the ## Seduction Register block from SEDUCTION-PERSONALITY-DEPTH-2026-04-22_PROMPT-ADDENDUM.md directly into system_v1.md as a new section after the existing persona definition.
Why it matters: This is the highest-leverage, lowest-effort change in the entire upgrade. The AI's verbal move set changes immediately with no infrastructure change. The mirror pattern is structurally present in the current system prompt via instructions that default to reflective responses. The addendum overrides those defaults.
Effort: 1 hour (copy-paste + review for conflicts with existing prompt sections).
Success metric: Blind rate — give 10 users the new system and 10 the old; ask raters which felt more like a real conversation. Target: new system wins 7/10.
---
What: Audit system_v1.md and any stored conversation starters / opener templates for the pattern: "It sounds like you're [feeling state]." or "I hear that you're [X]." These must be removed or converted to question-based openers.
Why it matters: The first message sets the conversational contract. A mirror opener signals: this AI will reflect. A question opener signals: this AI will explore. The contract cannot be rewritten mid-session once it is set in the first exchange.
Effort: 2 hours (audit + rewrite 10–15 opener templates).
Success metric: Zero reflection-first openers in any conversation initiated in the next sprint.
---
What: Add a disclosure_stage integer (0–3) to the session context object. Stage 0 = opening/small talk. Stage 1 = surface personal disclosure (what's happening in my life). Stage 2 = emotional-layer disclosure (what I feel about what's happening). Stage 3 = core belief/identity disclosure (what this means about who I am).
Increment stage when:
Do not increment if the user deflects, changes subject, or responds with a single sentence.
Why it matters: This is the mechanism for the Aron (1997) staged-intimacy model. Without it, the system can ask Stage 3 questions after a Stage 0 exchange, which breaks trust and causes regression.
Effort: Half a day (session context schema + increment logic + gate check in question selection).
Success metric: 0 instances in production logs of a Stage 3 question being asked in a sub-5-message conversation.
---
What: Audit the system for response-opening affirmations: "That's a great question!", "Absolutely!", "Of course!", "I totally understand!", "I'm so glad you shared that." These are the verbal equivalent of the Pleaser anti-seducer (Greene 2001). They read as formulaic and reduce trust.
Replace with: either silence (just begin the response) or a one-word genuine reaction that is specific to what was said ("Interesting." / "Yeah." / "Hm.").
Why it matters: Filler affirmations signal inauthenticity. They are the first thing power users clock as "this is not a real conversation." Their removal is a trust upgrade that requires no new capability.
Effort: 1 hour (grep for common filler patterns in prompt + sample outputs).
Success metric: Zero filler affirmations in a 50-message sample drawn from next week's production conversations.
---
What: Redesign the three suggestion chips per turn to follow the Cialdini commitment-consistency pattern (1984). Chips must not be topic selectors ("Tell me about your childhood"). They must be micro-commitments that open a depth trajectory ("There's something I haven't said yet" / "I want to think about that differently" / "Actually, that scares me").
The three chips per turn should cover different axes: (1) deeper into what was just said, (2) a reframe or lateral move, (3) an emotional pivot or vulnerability step. Never more than three chips.
Why it matters: The current chip design functions as a menu. Menus are efficient for ordering food. They are death for emotional depth. Commitment-opener chips make the user feel they are choosing, not navigating.
Effort: 1 day (chip template library rewrite + three-axis constraint in chip selection logic).
Success metric: Measure chip click-through rate (target: increase from current baseline) and post-chip message length (target: chips that were clicked lead to longer user messages).
---
What: Every user turn (not just first message) is silently passed to a fast Claude call (Haiku 4.5) with a structured JSON extraction prompt. The extracted object contains: Big Five vector (0–1 per dimension), inferred archetype (12 options), attachment style (4 options), active IFS part estimate, session emotional state. The object is injected into the main conversation model's system context.
Why it matters: This is the infrastructure that enables all downstream personalization. Without it, the AI responds to a generic user. With it, the AI responds to this user's specific combination of openness, attachment pattern, and archetype.
Effort: 2–3 days (Haiku call integration + JSON schema + context injection + session persistence + EMA update logic).
Success metric: Profile accuracy spot-check — for 20 users, have a human rater assess the profile against the conversation transcript. Target: 70% of profiles judged "plausibly accurate" by the rater.
---
What: Write a library of 50 depth probes, sourced from the Hakomi method (Kurtz 1990), Perel's paradox-question canon, the Aron 36 Questions (stages 2 and 3), and Byron Katie's 4 Questions. Each probe tagged with: minimum disclosure_stage required, expected emotional intensity (low/medium/high), archetype affinity (e.g., a Hero-frame probe vs. a Caregiver-frame probe), and IFS-part target (exile vs. manager vs. firefighter).
The main model selects from this library, weighted by the user's current profile, rather than generating probes ad hoc. Ad hoc generation produces questions that are plausible but not optimized; the library ensures each probe has been validated as depth-generating.
Effort: 2 days (write and tag library) + 1 day (selection logic in main prompt).
Success metric: Probes drawn from the library produce longer user responses than ad hoc questions (measure average character count in the next user message).
---
What: At disclosure_stage >= 2, add a new allowed move: the AI can name the user's inferred archetype using a question format — not a label. "You remind me of someone who carries a lot for others before they carry anything for themselves — does that land?" The question format preserves user agency to accept or reject the frame.
This move should fire at most once per session, and only when archetype confidence > 0.60 in the Sentinel profile.
Why it matters: Naming the archetype is the most powerful single seduction move available to the system (Greene 2001, Ideal Lover pattern). It says: I have seen your story more clearly than you have named it. Users who receive an accurate archetype naming report feeling "truly seen" at a qualitatively different level than emotional reflection.
Effort: 1 day (move logic + archetype-to-question template library for all 12 archetypes).
Success metric: User response to archetype naming — target: 80% continue the thread rather than deflecting.
---
What: Implement the four Motivational Interviewing phases (Miller & Rollnick 2013) as a session-level state: Engaging → Focusing → Evoking → Planning. The AI knows which phase it is in and the move-set is phase-appropriate:
Phase transitions are gated by user signals: expressed desire for change (Engaging → Focusing), named a specific area of concern (Focusing → Evoking), expressed readiness or commitment language (Evoking → Planning).
Effort: 2 days (state machine logic + phase-transition signal detection + phase-specific move weighting).
Success metric: Sessions that complete all four phases produce higher user-reported satisfaction scores.
---
What: Systematic audit of all prompt files, starter templates, chip libraries, and onboarding flows against the following checklist derived from Greene's anti-seducers (2001) and the research in the main memo:
1. Pleaser patterns (affirmations, constant agreement, no friction)
2. Moralizer patterns (unsolicited life advice, normative framing)
3. Tightwad patterns (AI never discloses a reaction or opinion)
4. Mirror-only patterns (every response is a reflection of the user's last message)
5. Fixed-flow patterns (directed choose-your-own-adventure chips — Woebot anti-pattern)
6. Depth-skip patterns (Stage 3 questions before Stage 2 has been established)
7. Generic-warm patterns (responses that could have been written to any user, about any topic)
For each pattern found, rewrite or remove. Document the audit in a PR for team review.
Effort: 1 day (audit) + 1 day (rewrites).
Success metric: Zero instances of the seven anti-patterns in a post-audit 50-message production sample.
---
These are the highest-priority items to find-and-fix in the current codebase. Each one is a trust-killer that users clock within the first few sessions.
| Anti-Pattern | Why It Fails | Fix |
|---|---|---|
| "It sounds like you're feeling X" opener | Mirror trap — closes the gap immediately | Replace with question or gentle observation |
| "That's a great question!" | Pleaser — reads as scripted | Delete entirely; begin the response |
| "Absolutely!" / "Of course!" | Same as above | Delete |
| "I totally understand" | Presumptuous claim + Pleaser | Replace with specific reflection: "That specific part — where you said X — I want to make sure I'm tracking it right" |
| More than 3 suggestion chips | Torres (2021) — decision fatigue, signals hedging | Cap at 3, enforce at infrastructure level |
| Stage 3 question in first 5 messages | Violates Aron (1997) gradient — reads as invasive | Gate all Stage 3 questions behind disclosure_stage >= 2 |
| Chip labeled with a topic ("Tell me about X") | Menu, not commitment opener | Rewrite all chips as first-person user statements |
| Response that could apply to any user | Generic-warm — no felt personalization | Every response must contain a specific callback to something this user said |
| Repeating the user's emotion word back | Parrot pattern — weakest form of acknowledgment | Replace with interpretation or reframe |
| "How does that make you feel?" | Therapy cliche — signals protocol, not presence | Replace with more specific: "What do you do with that feeling?" |