Work log: Phantasmagoria — March 22, 2026

What shipped today

Today was the biggest architectural shift in the project’s history: the complete narrative-first generation pipeline is now working end-to-end for all three event types — standalones, anomalies, and dig sites. The old roller-first approach (deterministic effects → AI writes text to match) has been replaced with a three-phase system: Phase 1 (AI writes the story freely), Phase 2 (AI assigns effects that match the narrative), Phase 3 (deterministic assembly into render-ready YAML).

The prompt engineering work was the core of the session. Each event type got its own Phase 1 template (PROMPT_STANDALONE.md, PROMPT_ANOMALY.md, PROMPT_DIG_SITE.md) with type-specific guidance — standalones get choice architecture patterns (all-good, all-bad, risk-vs-reward, gain-vs-loss), anomalies get survey context and planet-scoped deposit support, dig sites get multi-chapter arc planning with auto-resolve middle chapters. Phase 2’s outcome prompt was updated with choice pattern awareness and single-option follow-up support. Follow-up events are now defined inline by Phase 1 as resolution popups — single “Acknowledged” dismiss button, no new decision points.

Infrastructure improvements landed too: all AI model references moved from hardcoded strings to .env (AI_MODEL=claude-sonnet-4-6), release YAML can now override tone and reward_scope probability weights, and a new event_assembler.py helper centralizes Phase 3 assembly for all event types. Test harnesses exist for each type (test_narrative_first.py, test_narrative_anomaly.py, test_narrative_site.py) producing render-ready YAML with working follow-up wiring.

Completed

• #211 (partial) — Narrative-first architecture: Phase 1, 2, and 3 working for all event types
• #222 — Created and triaged: “Allow config for early/mid/late game timing in release YAML” (backlog)

Carry-over

• Integrate narrative-first prompts into the actual generator scripts — the test harnesses work end-to-end but generate_standalone.py, generate_anomaly.py, and generate_site.py still use the old prompt flow. The new prompts and assembler need to replace the old code paths.
• Phase 2 still uses society_research/physics_research as resource names — banned in the prompt but the AI occasionally ignores it. Needs stronger enforcement or post-processing cleanup.
• Tech variety is narrow — Phase 2 leans heavily on tech_mine_rare_crystals and tech_shields_3. The tech pool in outcome_prompt.py has more options but the AI gravitates to a few.
• Title diversity — with the celestial_equinox theme, the AI keeps generating “The Cartographer’s [X]” titles. The existing_events context will help in batch mode, but may need explicit “avoid these words” guidance.

Risks

• The Sonnet 4.6 model ID (claude-sonnet-4-6) works today but is an alias — if Anthropic changes the routing, all generation breaks. The .env approach means a one-line fix but we’d need to notice.
• Middle dig site chapters sometimes include options when they shouldn’t. The assembler handles this gracefully (treats 1-option chapters as auto-resolve) but it means wasted Phase 2 API calls in batch generation.

Flags and watch-outs

• PRODUCT.md principle #1 says “AI for creativity, determinism for balance” — the narrative-first approach has AI assigning effects (Phase 2), which shifts this principle. It’s working well but the product doc should be updated to reflect the new architecture.
• The old followup_generator.py still exists with the roller-first follow-up approach. It’s used by the existing generators but will be dead code once the new pipeline is integrated.

Next session

1. Integrate new prompts into real generators — replace the old prompt building in generate_standalone.py with PROMPT_STANDALONE.md + event_assembler.py. Same for anomaly and site. This is the bridge from “test harness works” to “batch generation uses the new flow.”
2. Generate a test release with the new pipeline — run a full celestial_equinox generation (3 standalones, 3 anomalies, 3 dig sites) and playtest it.
3. Clean up dead code — once generators use the new flow, remove old prompt building, old follow-up generator’s roller-based Phase 2, and deprecated functions.