Pipeline 4 · Candidate deep dives

Generates a Heat-style 7-section design document for a single candidate game.

python scripts/deep_dive.py list-candidates           # top candidates without a deep dive
python scripts/deep_dive.py prepare "Hot Streak"      # emits a self-contained agent prompt
# (run the agent against the prompt; output goes to notes/<slug>_deep_dive.md)

prepare doesn't call an LLM — it builds the prompt by loading:

• the BGG row from SQLite,
• prior v2/v3 enrichment if present,
• a calibration pointer to notes/heat_deep_dive.md (the gold-standard reference).

The agent then runs the prompt, and its output is checked into notes/<slug>_deep_dive.md.

The 7 sections

Every deep dive follows the same shape:

1. One-liner — the elevator pitch
2. Core loop — the minute-to-minute play
3. Decision space — what the player thinks about
4. Skill expression — what improves with mastery
5. Tactile dependency — how much the physical/touch layer carries the experience
6. Translation difficulty — what's hard about porting it
7. Live-service / meta-layer ideas — what an upgrade structure could look like

Currently published

8 deep dives are checked in. See /deep-dives for the list.