# NPCs — a capture pipeline, not a library **Revised 2026-08-24.** This directory originally proposed a pre-built NPC library. That was wrong, and the reasoning is worth keeping. --- ## The reversal The AI can generate a tavern keeper on demand, easily and well. Generation was never the problem. **What it cannot do is generate the *same* tavern keeper in session 9 that it improvised in session 3.** That is a **persistence** problem, not a generation problem — and session logs already solve persistence. A thousand pre-written people is an expensive fix for a problem we don't have. --- ## What we build instead ### 1. Generate in situ NPCs appear as the plot needs them. No pre-population, no roster to maintain, no directory of strangers nobody will ever meet. ### 2. Persist on contact **The moment a player interacts with an NPC, that NPC is written to canon.** This is the load-bearing mechanic. Everything else is convenience. Captured on first contact: ```yaml name: role: location: # bound at capture first_seen: # session ref — provenance stat_block: # SRD reference, only if combat is plausible wants: # ONE line voice: # how they actually talk, 1-2 lines tell: # a physical habit to hang narration on knows: # visibility-tagged — see architecture §3 - fact: visibility: # true | known | suspected | false ``` `knows` **must** carry visibility, or an NPC cannot hold a secret without the DM leaking it. See [`../docs/architecture.md`](../docs/architecture.md) §3. ### 3. Promote when remembered An NPC moves to `named/` the first time the player remembers their name unasked. That's the moment they stop being furniture and start being a character. ### 4. Seed generation from trait tables — the ONE thing a corpus buys Improvised NPCs drift toward a single voice. Ask for a blacksmith, then a guard, then a priest, and eventually the whole world sounds like one person doing voices. The fix is **variety at the point of generation**, not coverage in advance. [mjmcphee/dnd-npc-generator](https://github.com/mjmcphee/dnd-npc-generator) by Mike McPhee is MIT and genuinely **2024-native** (uses "species", not "race"). ⚠️ **Right-sized 2026-08-24:** it is a **small personal project** — 1 star, 1 fork, 4 commits — and its tables are **inline Python dictionaries inside the script**, not data files. It is a **starting point, not a corpus.** Expect to extend it heavily, and don't plan around it as though the problem is solved. See [`../docs/open-source-resources.md`](../docs/open-source-resources.md) §4. --- ## What we DO pre-build The default is **only what must be mechanically correct** — stat blocks, which already live in [`../library/`](../library/) as SRD data. The split is clean: | | pre-built | generated | |---|---|---| | **stat block** — what it does on its turn | ✅ correctness matters | | | **personality** — what it wants, how it talks | | ✅ variety matters | Correctness gets retrieved. Character gets invented, then *captured*. ### Two sanctioned exceptions **Revised 2026-08-24.** Both are pre-built *people*, not stat blocks, and both earn it for the same reason: **generation cannot produce them at the moment they are needed.** | | why it is exempt | |---|---| | [`adventurers/`](adventurers/) | Rival parties and hirelings are built like characters rather than like monsters. Six, converted from superseded player sheets. | | [`antagonists/`](antagonists/) | **A nemesis is defined by recurrence**, and recurrence cannot be improvised — the accumulation has to have been designed before the first meeting. Six, and six is the ceiling. | Nothing else here is pre-built, and the bar for adding a third exception is the same test: *does improvising this at the table produce the wrong thing, rather than merely a different one?* --- ## Structure ``` npcs/ tables/ roles/ town/ 20 roles — people who live in a settlement world/ 12 roles — people who live outside one hostile/ 10 command-tier roles — droppable threats names/ traits/ hooks/ generation seeds (⬜ unbuilt — VLUM task 2) adventurers/ 6 rival adventurers, built like characters antagonists/ 6 recurring nemeses, built like characters ``` ⚠️ **Captured NPCs are not stored here.** They live inside the campaign, at `LIVE//canon/locations//npcs/`, because **canon is campaign-bound**. There is no `captured/` and no `named/` directory — promotion is a `status:` field, so a path never changes. See [`../docs/npc-containers.md`](../docs/npc-containers.md) §1 and [`../docs/campaign-start.md`](../docs/campaign-start.md) §1. --- ## Status | piece | state | |---|---| | `tables/roles/` | ✅ **42 built** — 20 town, 12 world, 10 hostile | | `adventurers/` | ✅ **6 built** | | `antagonists/` | ✅ **6 built 2026-08-24** | | `tables/names/ traits/ hooks/` | ⬜ **the remaining gap.** Mechanical, unblocked, no design required. | The capture pipeline itself depends on the canon store (build step 5).