diff --git a/skills/rpg/SKILL.md b/skills/rpg/SKILL.md
new file mode 100644
index 0000000..2df2d10
--- /dev/null
+++ b/skills/rpg/SKILL.md
@@ -0,0 +1,122 @@
+---
+name: rpg
+description: >-
+  Tim's tabletop RPG scenario-writing system — house structure, GM-readable
+  writing style, pregen design, playtesting protocol, production pipeline
+  (handouts, art, GM console), and memory/repo integration. Use this skill
+  whenever the work touches tabletop RPG scenarios or adventures in ANY way:
+  writing or revising a scenario, one-shot, con game, adventure, module,
+  campaign, session, or BPN; designing pregens, NPCs, clues, handouts, chases,
+  or set-pieces; playtesting; preparing convention slots (Continuum,
+  Contingency, Airecon, Convulsion, ChaosiumCon); or working with systems like
+  Blade Runner, Call of Cthulhu, SLA Industries, Mothership, Traveller,
+  Pendragon, Daggerheart, Shadowdark, Stormbringer, RingWorld, BRP, Delta
+  Green. Trigger even if the user doesn't say "scenario" — "prep my Friday
+  slot", "new adventure for X", "fix the pregens", "make handouts" all count.
+---
+
+# RPG Scenario Writing
+
+Tim writes tabletop RPG scenarios — mostly 4-hour convention one-shots, sometimes
+longer campaign material — and runs them as GM. This skill is the house system:
+structure, style, production pipeline, and the memory loop that keeps it all alive
+across machines and sessions.
+
+## The prime directive: write for the GM at the table
+
+A scenario is a performance document, not a novel. The GM reads it in 30-second
+glances while four people wait. Every formatting decision serves that:
+
+- **Scannable over literary.** Bullets, tables, bold keywords. Prose only where it
+  earns its place (read-aloud text, plot summary, atmosphere).
+- **Information in the order the GM needs it**, not the order the story happens.
+- **Read-aloud text in `>` blockquotes**, ≤3 sentences by default (players
+  disengage after two; only the cold open may run to 4–6), always labelled
+  "read aloud or paraphrase". Longer material becomes bullets the GM voices.
+- **Mechanics inline, bracketed:** `[BR: Observation — spots the torn hem]`.
+  Never send the GM to another page for a roll.
+- **Every NPC gets a one-line essence** (who they are + what they want) before
+  any detail, and appears again in the end-of-document **NPC Roster quick sheet**.
+- **State the failure case.** Every clue, scene, and set-piece says what happens
+  if players miss it, skip it, or break it.
+
+## Workflow
+
+### 1. Orient (start of any RPG task)
+
+- Query Graphiti (`search_memory_facts`) for the scenario/convention/system at hand.
+- Find the working directory. Convention material lives under
+  `~/Library/Mobile Documents/com~apple~CloudDocs/RPGS/Conventions/<Event Year>/...`
+  and is mirrored to a private GitHub repo per event (e.g. `slaguru666/Continuum2026`).
+- Ask which **scenario type** this is only if it isn't obvious:
+  - **Convention one-shot** (default) → read `references/convention-one-shot.md`
+  - **Campaign / multi-session** → read `references/campaign.md`
+- For craft questions (structure, clues, pacing, mystery design) read
+  `references/craft.md`. Before writing or revising anything, skim
+  `references/lessons.md` — it is the accumulated playtest wisdom.
+
+### 2. Build
+
+Order of work for a new scenario:
+
+1. **Concept + Design intent** — 3–5 numbered statements of what this scenario
+   is and deliberately is not. Include an elevator pitch line (the tagline under
+   the title) and content warnings.
+2. **Plot Summary (GM truth)** — the full hidden truth, then a **timeline** of
+   events before play begins. Write this before any scenes: scenes are windows
+   onto a situation that already exists.
+3. **Structure pass** — acts, timing table with GM checkpoints, countdown clock
+   of what antagonists do on schedule.
+4. **Scenes, NPCs, clue trails** — per act, using `assets/scenario-template.md`.
+5. **Pregens** — read `references/pregens.md`. Wire each PC into the case.
+6. **Production** — handouts, art, GM console, print: `references/production.md`.
+7. **Playtest** — read `references/playtest.md`. A scenario is not
+   "convention-ready" until it has survived desk playtests and the fix list is
+   applied and versioned.
+
+### 3. Close the loop (end of any RPG session)
+
+This skill is **live** — it must get smarter every time it is used:
+
+- **Graphiti:** `add_memory` with decisions made, scenario state, and anything
+  learned (structure fixes, ruling patterns, player-behaviour discoveries).
+- **Repo:** commit and push scenario work to its event repo.
+- **Skill self-update:** if the session produced a transferable lesson (not
+  scenario-specific detail), append it to `references/lessons.md` with a date
+  and one-line context, then sync the skill across machines:
+  `cd ~/Git/claude-config && ./sync.sh "rpg skill: <what changed>"`.
+  Scenario-specific facts go to Graphiti, *transferable craft* goes to lessons.md
+  — that distinction keeps the skill lean.
+
+## House conventions (all scenario types)
+
+- **Files are Markdown**, one scenario per file, self-contained (a GM with only
+  that file can run the game). HTML only for print artifacts (character sheets,
+  handout packs, GM console).
+- **Folder shape per scenario:** `scenarios/` `characters/` `handouts/`
+  `reference/` `playtest/` `gm-utility/` `print/` + `README.md` index with
+  status line (version, playtest count, remaining prep).
+- **Versioning:** v1 draft → playtest → v2, v2.1... Status becomes
+  "convention-ready" only after a full-session playtest lands inside the time
+  budget. README tracks version + what remains.
+- **Mechanic tags:** `[SYSTEM: Skill — effect]` e.g. `[CoC: Spot Hidden — ...]`,
+  `[BR: Manipulation — ...]`. Include a 3–4 line system crash-course in the
+  Master Rules block so a stand-in GM could run it.
+- **Art:** ink sketch / pencil illustration, hand-drawn feel, never
+  photorealistic. Asset IDs `[Slot]-ART-[##]`, handouts `[Slot]-H[act]-[letter]`.
+  Generate via `mj-gen "<prompt>"` (Midjourney bridge). Each scenario defines
+  its own art direction line on top of the base style.
+- **Licence footer** on READMEs: private convention play materials, not for sale.
+
+## Reference files
+
+| File | Read when |
+|---|---|
+| `references/convention-one-shot.md` | Building/revising a 4-hour slot scenario (the default) |
+| `references/campaign.md` | Multi-session or open-ended material |
+| `references/craft.md` | Structure, clue, pacing, mystery design decisions |
+| `references/pregens.md` | Creating or auditing pre-generated characters |
+| `references/playtest.md` | Running desk playtests, writing fix lists |
+| `references/production.md` | Handouts, art, GM console, print, repo layout |
+| `references/lessons.md` | Always skim before writing; append new lessons |
+| `assets/scenario-template.md` | Starting a new scenario document |
diff --git a/skills/rpg/assets/scenario-template.md b/skills/rpg/assets/scenario-template.md
new file mode 100644
index 0000000..f3384c8
--- /dev/null
+++ b/skills/rpg/assets/scenario-template.md
@@ -0,0 +1,272 @@
+# [Event – Slot] – [System]: [TITLE]
+
+> *[Tagline — one line, the scenario's hook as a whisper.]*
+
+---
+
+## GM ESSENTIALS (one glance, any GM)
+
+- **System crash-course:** [3–4 lines on the core roll, so a stand-in GM could
+  run this. e.g. attribute die + skill die, 6+ succeeds, push to re-roll…]
+- **Mechanics inline:** `[SYSTEM: Skill — what it reveals or does]`
+- **The clock counts play time** — pause it for the break. All cumulative
+  times and checkpoints below are on that clock.
+- **Hard rule: at 2:45, go to Act Three regardless.** Fast-tracks per act.
+- **Safety:** state tool in use (X-card / open door) during seating — 30
+  seconds, no ceremony.
+
+---
+
+## Overview
+
+- **Event:** [Convention — day, slot, time, venue]
+- **System:** [Game + publisher]
+- **Setting:** [Where and when]
+- **Tone:** [Two or three precise words, then one sentence]
+- **Maximum Runtime:** 3 hours 30 minutes play + 10 minute break
+- **Players:** [N] pregens (runs [N−1]–[N+1]; drop [PC] first at fewer; at
+  [N−1] the splits become [adjustment])
+- **Content warnings:** [List — conventions require them, players deserve them]
+- **Art Style:** Ink sketch / pencil illustration. [Scenario-specific direction line]
+
+### Design intent (read this first)
+1. [What kind of scenario this is — e.g. investigation-first, no required combat]
+2. [Structural promise — e.g. built to split into 2+2 parallel tracks]
+3. [Ending shape — closure or designed ambiguity, and why]
+4. [What this scenario deliberately is *not*]
+
+---
+
+## Session Timing Guide
+
+| Section | Duration | Cumulative (play) | GM Checkpoint |
+|---------|----------|-------------------|---------------|
+| Slot Zero + Introduction | ~15 min | 0:15 | [One-line test the GM can apply] |
+| Act One — [Title] | ~60 min | 1:15 | [Essential clue found; hook to Act Two live] |
+| *Break — pause the clock* | *10 min* | *1:15* | *Lands on the Act One cliffhanger* |
+| Act Two — [Title] | ~90 min | 2:45 | [Truth revealed; heading to confrontation] |
+| Act Three — [Title] | ~30 min | 3:15 | [Confrontation resolved] |
+| Epilogue | ~15 min | 3:30 | [Vignettes delivered, debrief done] |
+
+Durations are **ceilings, not fills** — plan ~3:15 of content against them.
+**Hard rule:** at 2:45, go to Act Three regardless.
+
+---
+
+## Plot Summary (GM truth — players never see this page)
+
+*The full hidden truth, start to finish, regardless of player action. What
+actually happened, who wants what, what the antagonist does next.*
+
+**Timeline (GM reference):**
+- **T-[n]:** [event before play]
+- **Game night:** [where play begins]
+- **[Deadline]:** [the convergence point]
+
+### Revelations (GM checklist — 8–12, abstracted from where they're found)
+
+| # | Revelation | Expected by | Delivered? |
+|---|-----------|-------------|------------|
+| R1 | | Act One | |
+| R2 | | Act Two | |
+
+*Anything undelivered by end of Act Two: fold into Act Three's confirmation
+scene or consciously drop — the Act Two Pacing Note says which.*
+
+---
+
+## Player Experience
+
+*One paragraph, second person: the scenario as the table will live it. This is
+the pitch — if it reads flat, the scenario is flat.*
+
+---
+
+## The Countdown (background clock)
+
+*Antagonists move on schedule whether or not the players are watching. When the
+party stalls, advance one step and let the players feel it. Opposition stays
+one step behind the players until the climax. This table is also the universal
+fallback: any missed essential clue can arrive as a Countdown event.*
+
+| Clock | Event |
+|---|---|
+| [t0] | [Opening state] |
+| [t1] | [Antagonist move] |
+| [deadline] | [Convergence] |
+
+---
+
+## Pre-Game Handouts
+
+| ID | Description | Format | When Given |
+|----|-------------|--------|------------|
+| [Slot]-H00-A | | | |
+
+---
+
+## Introduction — [Title] *(~15 minutes, including Slot Zero)*
+
+### Slot Zero (before the fiction)
+1. Seating + safety line (30 seconds)
+2. Pitch each pregen aloud in one tagline (the sheet's italic essence line)
+3. Players choose — or GM deals — in ≤5 minutes
+4. Open in motion; introductions surface during the action
+
+### Opening (read aloud or paraphrase)
+
+> [≤3 sentences default; the cold open may run to 4–6. Sensory, in motion,
+> ends pointing at the players. Longer material becomes bullets the GM voices.]
+
+### Player Briefing
+- [What the players know, as bullets — victim, job, constraint, authority]
+- [Which handout to give, and when]
+
+### Atmosphere Notes
+- [Sound, light, smell — the loops the GM returns to]
+- [Suggested audio — home tables only; con halls are loud]
+
+### Hook into Act One
+[The choice or thread that moves them — offer, don't assign.]
+
+---
+
+## ACT ONE — [Title] *(~60 minutes)*
+> **Goal:** [What this act establishes; what must be true before Act Two.]
+
+### Summary
+[What happens here, in five lines or fewer.]
+
+### Split Plan *(if this act runs as parallel tracks)*
+- **Recommended pairing:** [2+2 by PC lanes — offer, don't assign]
+- **Segments:** alternate ~12 min, end each on a pull line: "[e.g. what was the hurry?]"
+- **Cross-team payoff:** [what each half learns that the other needs]
+
+### Locations
+
+#### [Location Name] (Track A/B if split)
+- **First impression:** [what they see/hear/smell on arrival]
+- **Interactables:** [what rewards attention]
+- **Hidden:** [optional/extra material only — essential clues never live here]
+  `[SYS: Skill — effect]`
+- **Artwork:** [Slot]-ART-[##]
+
+### NPCs
+
+#### [NPC Name]
+*[One-line essence: who they are + what they want.]*
+- **Physical:** [one evocative detail]
+- **Motivation:** | **What they know:** | **Disposition:**
+- **Mechanics:** [system stat line, inline — no page refs]
+
+### Clue Trail — Act One
+
+| # | Clue | Location / Source | How Found | What It Reveals | Type |
+|---|------|-------------------|-----------|-----------------|------|
+| 1 | | | | | Essential |
+| 2 | | | | | Optional |
+
+**Fallbacks (one per essential clue):** [How the GM introduces each naturally
+if missed — NPC volunteers it, environmental prompt, or Countdown event.]
+
+### Story Hooks (if players stall)
+- [Proactive event — the world moves]
+
+### Handouts
+
+| ID | Description | Format |
+|----|-------------|--------|
+| [Slot]-H01-A | | |
+
+### Pacing Note
+[What must be true before Act Two. At 1:15:]
+> Fast-track: [specific action to jump forward]
+
+---
+
+## ACT TWO — [Title] *(~90 minutes)*
+> **Goal:** Investigation, revelation, escalation. The players make the choices that define the ending.
+
+[Same section pattern as Act One — Split Plan, Locations, NPCs, Clue Trail
+(continue the numbering), Story Hooks, Handouts. Additionally:]
+
+### [THE SET-PIECE] (one per scenario)
+- [Round-by-round or beat-by-beat GM structure; own quick-ref page in `reference/`]
+- **Idle players:** [the job the sidelined pair runs — obstacle cards, bystanders]
+- **If it breaks:** [escape, refusal, early violence — what then]
+
+### Pacing Note
+At 2:45, move to Act Three regardless. [Name here which undelivered
+revelations fold into Act Three and which get dropped.]
+> Fast-track: [specific jump to the climax]
+
+---
+
+## ACT THREE — [Title] *(~30 minutes)*
+> **Goal:** Consequence. Fast pace, high stakes. No new information — only confirmation.
+
+### Opening (read aloud)
+> [Short. Arrival at the convergence.]
+
+### The Situation
+[Who is present, what they want, what the clock says.]
+
+### The Choices (present plainly; then get out of the way)
+- [Choice A — and its cost]
+- [Choice B — and its cost]
+- [Compromise options playtests or tables have found]
+
+### If Someone Shoots
+[The chaos contingency — the table that torches the negotiation gets a real ending too.]
+
+### Story Hooks (contingencies)
+[If players freeze, or head for a catastrophic ending.]
+
+### Pacing Note
+30 minutes maximum. If unresolved at 3:15:
+> [The pressure line — deploy at the *first* sign of circular debate.]
+
+---
+
+## Epilogue — [Title] *(~15 minutes)*
+
+### Outcomes
+
+| Outcome | Trigger Condition | What Happens |
+|---------|------------------|--------------|
+| **[Named ending 1]** | | |
+| **[Named ending 2]** | | |
+| **[Failure / Pyrrhic]** | | |
+
+### Closing Scene (read aloud, adapt to outcome)
+> [Land the last line on the theme. Per-PC vignettes if earned.]
+
+**Early finish:** expand the vignettes, then run The Tangent as a bonus scene.
+
+---
+
+## NPC Roster Summary (GM quick sheet)
+
+| NPC | Essence | Wants | Key line |
+|---|---|---|---|
+
+---
+
+## Artwork Assets
+
+- **Art Style:** Ink sketch / pencil illustration. [Direction line.]
+
+| Asset ID | Description | Used In | Generation Prompt |
+|----------|-------------|---------|-------------------|
+| [Slot]-ART-01 | Cover | Intro | |
+
+---
+
+## The Tangent
+*One alternative direction the story could have taken, or the thread a future
+session would pull. Doubles as the early-finish bonus scene.*
+
+---
+
+## GM Notes (private)
+*Pacing advice, tone reminders, staging instructions, things to watch for.*
diff --git a/skills/rpg/references/campaign.md b/skills/rpg/references/campaign.md
new file mode 100644
index 0000000..f80de35
--- /dev/null
+++ b/skills/rpg/references/campaign.md
@@ -0,0 +1,50 @@
+# Campaign & Multi-Session Material
+
+Tim will say when material is campaign-shaped rather than a convention one-shot.
+The GM-readability rules, clue discipline, and production standards all still
+apply; what changes is the shape of time.
+
+## What changes from the one-shot format
+
+- **Timing loosens.** No 3h30 hard ceiling, no 2:45 rule. Keep per-session
+  targets and checkpoints, but sessions may end mid-arc.
+- **Endings invert.** Sessions *should* end on unresolved peril, an escalating
+  bang (stakes suddenly transformed), or an earned quiet character beat — the
+  cliffhanger types a one-shot must avoid. The *arc* still needs a written
+  destination.
+- **Hooks must respect choice.** Con players opted in by sitting down; campaign
+  players resent coerced hooks. Offer situation-hooks the party can approach
+  from any angle, and plant **long-term hooks** whose payoff is sessions away.
+- **Foreshadowing is a discipline.** Before writing session N, re-read what is
+  planned for N+1..N+3 and seed their elements now. Twists only land if the
+  clues predate them.
+- **Carry-forward bookkeeping** at the end of each session:
+  - Undiscovered revelations roll into the next session's list.
+  - Fronts advance a grim portent; clocks tick for everything ignored.
+  - Log NPC status, faction posture, promises made, and loose player theories
+    (player theories are free plot — steal the good ones).
+  - Persist all of this to Graphiti so any machine can pick the campaign up.
+- **Structure by fronts, not acts.** Bundle dangers into fronts, each with an
+  impending doom and escalating portents. Prep situations per the Alexandrian;
+  let acts emerge in play.
+- **Characters progress.** Advancement, gear drift, and relationship maps need
+  tracking; pregen audit rules from `pregens.md` apply to any handed-out PCs
+  but expect player-authored characters here.
+
+## Session prep loop (Sly Flourish's 8 steps, house-adapted)
+
+1. Review the characters (and last session's carry-forward log)
+2. Create a strong start — in motion within 20 minutes
+3. Outline potential scenes (brief, disposable)
+4. List secrets and clues — ~10 one-liners, abstracted from location
+5. Develop locations (immediate impression → interactables → hidden)
+6. Outline NPCs (one-line essence + want)
+7. Choose opposition
+8. Select rewards
+
+## Convention scenario ↔ campaign reuse
+
+A convention one-shot can seed a campaign: "The Tangent" section of every
+one-shot names the thread a future session would pull. Going the other way,
+compressing campaign material into a slot means cutting to one situation, one
+set-piece, one decision — and writing a real ending.
diff --git a/skills/rpg/references/convention-one-shot.md b/skills/rpg/references/convention-one-shot.md
new file mode 100644
index 0000000..b44d724
--- /dev/null
+++ b/skills/rpg/references/convention-one-shot.md
@@ -0,0 +1,121 @@
+# Convention One-Shot Structure
+
+The house format for 4-hour convention slots. This structure is proven across
+Continuum, Contingency, ChaosiumCon and Airecon games — deviate only with a
+stated reason in the Design intent.
+
+## Master Rules
+
+### Structure
+- **Introduction → Act One → Act Two → Act Three → Epilogue**
+- Maximum runtime **3h30 (210 min) of play** inside a 4-hour slot, plus a
+  10-minute break — ~3h40 wall clock, leaving real margin: strangers at a con
+  table run slower than any playtest.
+- Budgets: Intro 15 · Act One 60 · Act Two 90 · Act Three 30 · Epilogue 15.
+  These are **ceilings, not fills** — plan ~3:15 of actual content against
+  them. A scenario that playtests "on schedule with zero slack" is over-stuffed.
+- **The session clock counts play time.** Pause it during the break; every
+  cumulative time and checkpoint in the scenario is on that clock. (The GM
+  console's clock supports this.)
+- **Act Three always happens.** Compress Act Two if needed; never sacrifice the
+  climax. Hard rule: at 2:45 on the clock, go to Act Three regardless.
+- Every act ends with a **Pacing Note**: what must be true before the next act,
+  and a specific fast-track action if the clock has beaten the table.
+- Place the 10-minute **break** at the end of Act One, on a natural cliffhanger.
+- A one-shot must **end written**: closure or *designed* ambiguity, never an
+  unwritten "to be continued". The Epilogue includes an outcomes table
+  (Full / Partial / Failure / Pyrrhic — or scenario-specific named endings)
+  and an adaptable closing read-aloud.
+
+### Timing table
+Include a Session Timing Guide with a **GM Checkpoint** column — a one-line
+test the GM can apply mid-session ("both copies met, chase done, deadline
+known"). Checkpoints are what make the timing table usable in play.
+
+### Clues — the house standard
+Every essential clue has **three routes** to the players:
+1. **Automatic on the right approach** — the right skill or question in the
+   right scene finds it without a roll; rolls only buy *extra* information.
+2. **A written fallback** — NPC volunteers it, environmental prompt, second
+   location.
+3. **The Countdown** — a proactive antagonist move that delivers or forces the
+   revelation (this is the universal third route; it's why the Countdown table
+   is mandatory).
+
+Plus:
+- Clues form a **guided trail** — each points to the next, but players should
+  feel they discovered it, not followed a map.
+- Every act: at least one **essential** clue and one **optional** clue
+  (optional = reward for thorough play; trail survives without it).
+- Essential clues never live behind a search roll or in a location's "Hidden"
+  line — Hidden is for optional/extra material only.
+- Clue trail is a table per act: # | Clue | Location/Source | How Found |
+  What It Reveals | Type. Number clues continuously across acts, one fallback
+  line per essential clue.
+- Keep a **Revelations checklist** (8–12 one-liners, abstracted from where
+  they're found) after the Plot Summary. Anything undelivered by the end of
+  Act Two is folded into Act Three's confirmation scene or consciously
+  dropped — the Act Two Pacing Note says which. Beyond that, Act Three
+  contains **no new information** — only confirmation and consequence.
+
+### The countdown clock
+Antagonists and world events move on schedule whether or not players act.
+Write a timetable ("The Countdown") the GM advances when the party stalls —
+and make each advance *felt* (a glimpsed van, a witness gone quiet). Keep the
+opposition **one step behind the players until the climax** — pressure, not
+punishment.
+
+### Design intent block
+Right after the Overview, 3–5 numbered statements, e.g. "Investigation-first:
+no required combat", "Built to split: Acts One and Two run as 2+2 parallel
+tracks", "Ambiguous ending: conversation, not firefight". This is the contract
+the rest of the document must honour, and the first thing a reviewer checks.
+
+### Plot Summary + Player Experience
+Two views of the same scenario, both mandatory:
+- **Plot Summary (GM truth)** — what actually happened/is happening,
+  player-independent, with a dated timeline. Players never see this page.
+- **Player Experience** — one paragraph of the scenario as lived from the
+  table, second person, spoiler-shaped but honest. This is the pitch — if it
+  reads flat, the scenario is flat.
+
+## Slot logistics
+- Default **4 players** with pregens (6 max for some systems — pick per game
+  and say so). The Overview states the scaling line: which player counts the
+  scenario runs at, which pregen drops first at fewer, how splits adjust.
+  Nothing essential may live *only* on one PC.
+- **Safety tools, stated in 30 seconds, no ceremony** (X-card or open door —
+  UK cons expect it, and scenarios with private seams that cut close need the
+  exit ramp). One line in the Introduction covers it.
+- **Slot Zero** (the first 15 minutes, before the fiction): seating + safety
+  line → pitch each pregen in one spoken tagline (the sheet's italic essence
+  line) → players choose or GM deals in ≤5 minutes → open in motion.
+- **Party splits** are a feature: parallel scene tracks (Track A / Track B)
+  with engineered cross-team payoffs — each half holds half of a revelation.
+  Alternate between tables in ~12-minute segments, ending each segment on a
+  pull ("what was the hurry?").
+- One **set-piece** per scenario (chase, heist, séance...) with its own GM
+  quick-reference page. During set-pieces, give idle players jobs: hand them
+  the obstacle cards, let them voice the bystanders.
+- **Content warnings** in the Overview. Conventions require them and players
+  deserve them.
+
+## Endings
+- Aim the climax at a **decision, not a die roll**. The best Act Three is a
+  situation where every choice costs something and the scenario honours all
+  of them ("nobody gets to be right; the players decide who pays").
+- Write an "If someone shoots" (or equivalent chaos) contingency — the table
+  that torches the negotiation must also get a real ending.
+- List **compromise endings** players will invent (partial trades, third
+  options). If a playtest finds one, write it in.
+- Epilogue delivers per-PC **vignettes** keyed to the outcome, then table
+  debrief. Land the last line on the scenario's theme.
+- **Early finish** (table cracks it at 2:30): expand the per-PC vignettes,
+  then run The Tangent as a bonus scene — it's already written.
+
+## Fitting the slot (cut discipline)
+- Plan ~3:15 of content against the 3:30 budget; playtest timing includes
+  setup drift.
+- Pre-write the cuts: each act's Pacing Note names what to drop first.
+- Deploy pressure lines (an NPC checking a watch) at the *first* sign of
+  circular table debate, not when the clock is already lost.
diff --git a/skills/rpg/references/craft.md b/skills/rpg/references/craft.md
new file mode 100644
index 0000000..34bf1e8
--- /dev/null
+++ b/skills/rpg/references/craft.md
@@ -0,0 +1,88 @@
+# Scenario Craft
+
+Distilled from the best published design writing (The Alexandrian, GUMSHOE/Robin
+Laws, Sly Flourish, Sean McCoy's Mothership layout school, convention-GM guides
+by Katrina Ostrander and Burn After Running). These are the field's proven rules;
+the house structure in `convention-one-shot.md` already embodies most of them —
+use this file when making design *decisions* or debugging a draft that isn't
+working.
+
+## Situations, not plots (The Alexandrian)
+
+- Prep **situations** (locations, NPCs with goals and resources, available
+  information), never event chains — a scripted plot is a chain of potential
+  points of failure.
+- Write **goal-oriented opponents**: what the antagonist is trying to do, on
+  what timetable, with what tools (squads, safehouses, informants). Improvise
+  their reactions; never script them.
+- Always write the **adversary timeline**: what happens if the PCs do nothing.
+  (House form: The Countdown.)
+
+## Clues and mystery (Three Clue Rule + GUMSHOE)
+
+- The Alexandrian's Three Clue Rule (three independent routes to every
+  necessary conclusion) is implemented as the **house standard** in
+  `convention-one-shot.md`: automatic-on-right-approach + written fallback +
+  Countdown event. Apply that standard; this section explains the why.
+- **Never gate progress behind a die roll to notice something.** Finding core
+  clues is automatic for the right approach in the right scene; rolls buy
+  *extra* information. The game is *interpreting* clues, not finding them.
+- A **core clue** points to another scene. Test: if it's not worth writing in a
+  player's notes, it isn't a core clue. Write each scene with lead-ins and
+  lead-outs.
+- **Don't author red herrings** — players manufacture their own false leads for
+  free. Over-provision true information instead; the sorting is the gameplay.
+- Design **backwards from the truth**: write what actually happened and what
+  the villain does next, then lay the trail from the trigger event.
+- Keep a **revelation checklist** (the 8–12 things players must eventually
+  learn), abstracted from where they're discovered — any scene can deliver a
+  homeless revelation. Carry undiscovered ones forward. (Sly Flourish's
+  "secrets and clues".)
+- Unstick tools, in order: fallback delivery of the missed clue → **proactive
+  antagonist event** (a new death, the opposition strikes) → invent a new core
+  clue on the spot. Structures serve the table, not vice versa.
+
+## Pacing (convention wisdom)
+
+- **One substantive scene per hour of real play**; a 3h30 slot holds ~5. The
+  critical path should fill only about *half* the slot — deliberation, rules,
+  and jokes eat the rest. (Total planned content ≈ 3:15 against the 3:30
+  budget, per the house cut discipline.)
+- **Dice roll within the first 20 minutes.** Open in motion (in medias res or
+  pressure-adjacent) and let introductions surface during the action, not as
+  round-the-table monologues.
+- Premise formula for pitches: "When [disruptive event], will [adjective]
+  [group] manage to [verb] the [thing]?"
+- Introduce mechanics when first needed, not in a front-loaded rules lecture.
+- Spotlight is a resource: track who hasn't had a meaningful roll or scene and
+  aim the next beat at them.
+- Personal Agendas (Free League cinematic model): per-act secret goals on the
+  pregens are a drama engine — quiet tension in Acts One–Two, open
+  confrontation fuel in Act Three. The house "Private box" is the single-arc
+  version; use per-act agendas when the scenario needs PvP-adjacent heat.
+
+## Structural skeletons worth stealing
+
+- **Five Room Dungeon** (any genre): Guardian/Entrance → Puzzle or Social →
+  Trick/Setback → Climax → Revelation/Reward. Good emergency skeleton for a
+  short slot or a side episode.
+- **Fronts** and **progress clocks** (see `campaign.md`): the Countdown is the
+  wall-clock variant; use segment clocks when antagonist triggers are
+  event-driven rather than time-driven.
+
+## Writing for the running GM (McCoy / OSR layout school)
+
+- **The cascade:** can this prose be bullets? the bullets a table? the table a
+  diagram? Always compress downward.
+- Tell the GM **the truth on page one** — mystery is for players, never for the
+  reader. (House form: Plot Summary before any scene.)
+- Read-aloud text: players disengage after ~two sentences. Keep boxed text
+  short and sensory, or replace with a bullet list of details the GM voices in
+  their own words.
+- No mid-scene page-flips: stat lines live next to their encounter, maps next
+  to their key, cross-references carry explicit pointers.
+- Immediate impression first, then interactables, then hidden things — in every
+  location key.
+- Expressive terseness: telegram-minimal strips usable flavour; purple prose
+  buries the one fact that matters in paragraph three. One evocative detail per
+  element, load-bearing words in bold.
diff --git a/skills/rpg/references/lessons.md b/skills/rpg/references/lessons.md
new file mode 100644
index 0000000..efab0d3
--- /dev/null
+++ b/skills/rpg/references/lessons.md
@@ -0,0 +1,191 @@
+# Lessons Log
+
+Living file. Append transferable lessons here (dated, one-line context) at the
+end of any RPG work session; sync via `~/Git/claude-config/sync.sh`. Scenario-
+specific facts belong in Graphiti, not here — this file is only for lessons
+that will change how the *next* scenario is written.
+
+**Promotion rule:** when a lesson gets written into a reference file or the
+template, delete it from this log (git history keeps it) — otherwise this file
+grows into a duplicate of the skill. Several 2026-07 entries below are already
+encoded in the references; they stay as worked examples until the next pruning
+pass.
+
+## 2026-07 — AFTERIMAGE desk playtests (Blade Runner, Continuum 2026)
+
+- **Answer the obvious professional move.** Players will always try the
+  competent thing (field-scan the suspect, phone ahead to a contact, run a
+  city-wide facial search). If the scenario's premise hinges on such a move,
+  write its result explicitly — ideally as a sanctioned alternate route that
+  rejoins the trail, never a stonewall. Improvised rulings here are the
+  biggest holes playtests find.
+- **A player's key relationship is a loaded gun.** If a pregen's relationship
+  NPC guards a scenario target, decide in the text what happens when the
+  player leverages the relationship early. Done right it *improves* the scene
+  (the warned target is already spooked → better chase start).
+- **Formalise staging for twin/duplicate reveals.** When two groups hold
+  halves of a revelation, the NPC recites aloud while the GM hands the
+  matching card at the same moment. Never let a player's skim carry the scene.
+- **Idle pairs get jobs.** During any split set-piece, hand the sidelined
+  players the obstacle/bystander material to run. Formalise it in the
+  set-piece text.
+- **Deploy pressure lines early.** The written "NPC checks a watch" beat goes
+  in at the *first* sign of circular debate, not at the scheduled time.
+- **List the compromise endings.** Playtests and real tables invent partial
+  trades (wipe half the drive, surrender the copy). If a compromise is good,
+  it goes in the choices table — players will find it again.
+- **"On schedule with zero slack" is a fail.** Strangers run slower than any
+  desk test. Keep pre-written cuts per act and take them early.
+- **No-roll scenes need a sign.** If a beat is designed so no die roll
+  resolves it (a confession that stays ambiguous), say so in the text —
+  "no roll resolves this" — so the GM doesn't improvise one under pressure.
+
+## 2026-07 — Day One consistency audit (BRP, real-London setting)
+
+- **Real-city scenarios need a dedicated geography pass.** Check every compass
+  direction, "across the river", landmark sightline, and named route against
+  the actual map before calling a scenario ready — desk playtests won't catch
+  these, but any local player will (Day One had a hospital on the wrong bank,
+  a pier route through a tunnel that's past the destination, and a gatehouse
+  on the wrong end of the bridge).
+- **Grep the cast list for surname collisions between PCs and NPCs** — an
+  audit's cheapest catch (Kira Osei-Mensah vs Fatima Osei), and it reliably
+  confuses a convention table.
+- **In-fiction timestamps deserve arithmetic checks.** If the text says "two
+  hours before the alert", subtract the actual clock times; drafts drift as
+  timelines get re-plotted.
+
+## 2026-07 — Pregen roster rebuilds
+
+- **Audit inherited rosters against the current scenario, not just the
+  rules.** A mechanically-legal PC can still be skill-dead all night (a pilot
+  in a scenario with no vehicles). Rebuild the lane, keep the flavour, flag
+  the change for review.
+- **Signature items with a tiny mechanic** (once/session stress heal on
+  interaction) reliably produce unprompted table moments; pure-flavour items
+  go unused and nobody minds — both outcomes are fine, so always include one.
+
+## 2026-07 — Day One art production (mj-gen batch)
+
+- **Midjourney silently blocks banned words even in negation.** "no visible
+  gore" kills the job — the grid simply never arrives and the bridge times
+  out. Write the positive form instead ("understated, nothing graphic").
+- **Scene prompts drift photorealistic; style-lock them.** For the house ink
+  style, lead with "hand-drawn pen and ink illustration, NOT a photograph"
+  and append `--no photography, photorealism` — one dusk-boat prompt came
+  back as a photo without it.
+- **Maps are drawn, not generated.** MJ cannot do accurate labelled
+  geography; a hand-styled SVG (rough displacement filter, handwriting font,
+  cream paper) passes visually as sketch art and stays correct. Expect two
+  or three label-collision passes in a browser before it's clean.
+
+## 2026-07 — The Vain Crown build (house-system debut package)
+
+- **A rulebook's pregens are not con pregens.** VANITY's five shipped without
+  Vices — the system's own core engine. When packaging an included adventure
+  for a slot, audit the pregens against the system's *incentive* mechanics,
+  not just their stat legality.
+- **Debut systems need a dice primer baked into the Intro** — one staged
+  practice roll per player that doubles as the arrival scene, so teaching
+  costs zero extra minutes and the first Stumble happens somewhere safe.
+- **Give the villain a countdown table.** "Where is he now" per session-clock
+  row makes lost time (a failed pathfinding roll) mean something concrete and
+  lets fast tables genuinely change the finale's staging.
+- **Silence from a generation pipeline is a diagnosis, not a delay.** When a
+  batch goes quiet, check the actual channel/API for what the service said
+  before waiting again — and keep a fallback asset source (the system's own
+  existing art) so the package ships regardless.
+
+## 2026-07 — Princes Bride rebuild (Dee Sanction, Continuum 2026)
+
+- **Inherited multi-draft projects need a canon-reconciliation clause.** When
+  source materials disagree (two different charm-givers, two Eleanors, a
+  villain who is a fraud in one draft and a pragmatist in another), pick one
+  canon, write it as a numbered Design-Intent item ("Canon reconciled: ..."),
+  and list every superseded fact — otherwise the GM re-derives the
+  contradictions at the table.
+- **Verify the system against primary sources, not inherited quick-refs.** A
+  project's own GM reference had an entirely different game's dice mechanic
+  (PbtA 2d6 for a step-die Falter system). The character sheets + publisher's
+  site are ground truth; the quick-ref is a hypothesis.
+- **Run the set-piece on the system's own resolution mechanic.** The trial as
+  an open Verdict Die (the game's step-up/step-down, made visible on the
+  table) needed zero new rules and made evidence-presentation mechanical.
+  Find the system's native verb before inventing subsystems.
+- **Keep good inherited sheets; add an insert layer.** When existing pregens
+  are printed and well-made but generic, don't rebuild — print per-agent
+  mission inserts (private seam + dice crib + table-name conventions) that
+  slip inside the sheets. Cheaper, and the sheets stay canon.
+
+## 2026-07 — Silvery Moon rebuild (inherited-draft salvage)
+
+- **Verify a period scenario's real-world facts, then decide: correct or
+  canonize.** Silvery Moon's eclipse was astronomically impossible (new moon
+  that night — checked against the real 1892 record). Correcting it would
+  have cost the climax; canonizing the wrongness AS the horror (the moon that
+  should not be there, proven by an almanac handout and an astronomer pregen)
+  turned the draft's worst error into its best clue. Always check which move
+  the fiction wants before reaching for the eraser.
+- **Mine the draft's own margins for pregens.** The strongest pregen is often
+  already in the text as a named-but-unused NPC (Dr. Lyle was the draft's own
+  "recent collaborator"). Lift them before inventing.
+- **When two source layers contradict (prose vs app data), reconcile so both
+  are true** — Braddock "too afraid to act" AND "cult member" became a
+  coerced member, which was stronger than either. Contradictions are usually
+  two halves of a better fact.
+
+## 2026-07 — The Con Desk (multi-slot convention tooling)
+
+- **A convention weekend needs ONE entry point.** When several scenarios each
+  have their own GM console, build a hub with the con's real schedule baked
+  in: live-clock "up next / live now" badges, one-tap launches, and per-slot
+  pack-the-bag checklists. Every console gets a ⌂ link home. The hub's
+  relative links mean the whole repo folder must travel to the tablet
+  together — say so in the README.
+- **A contrarian pass earns its keep on the moves the GM prays nobody makes.**
+  "Cross the river NOW" and "we go back for the van" were both one-line hooks
+  until dice forced them; both became the best unwritten scenes. If a hook
+  can be driven to, it needs a written destination and its cost said aloud.
+
+## 2026-07 — Vain Crown big-table heroes (fixed-class-list systems)
+
+- **When the system has a closed class list, big-table pregens are variant
+  builds, not new classes.** Differentiate inside the class: a different
+  fighting style (sword-and-board Bram vs two-handed Hedda), a different
+  lane emphasis (stealth-killer Pinch vs thrown-blade showman Marlo), a
+  different creed/spell loadout (sustain Wren vs control grave-warden
+  Tobin). Show the audit math so the variants are provably legal.
+- **Point the new Vices at the existing roster.** "A Rival — [existing PC]"
+  only fires at player counts where both are dealt, which big-table extras
+  guarantee. Doctrinal Pride vs the other cleric's creed, and a Glory-Hound
+  pairing, gave every added seat built-in table chemistry for free.
+- **Scale the set-piece with the roster on the card.** The deal-order note
+  (6th/7th/8th) plus boss Grit per extra hero and mook multiplier lives ON
+  the GM cut-out corner of the last hero card page, not buried in the doc.
+
+## 2026-07 — RUNDOWN cardless chase system (Blade Runner, house subsystem)
+
+- **Do the expected-value arithmetic before committing to a subsystem.** A
+  spec self-review pass that actually computed the dice maths caught two dead
+  mechanics that read fine in prose: a role structure where the pursuers' net
+  was ~0.00 per round (the chase could literally never end), and a support role
+  worth +0.24 successes when simply doubling up on the main role was worth
+  +0.57 — nobody would ever have picked it. Prose hides both. One script found
+  both in a minute.
+- **A role that spends a resource needs a role that refunds it.** Four roles
+  only became an economy once the support role was recast from "adds successes"
+  to "buys off the cost the aggressive role generates". Test any new role menu
+  by asking what each option *trades*, not what it *does* — if two options
+  trade the same thing, one of them is decoration.
+- **Check whether the rules you're replacing already contain the escape
+  hatch.** The published Blade Runner chase ships a die table as an alternative
+  to its own obstacle deck, so "make it cardless" was already solved in the
+  book. The real brief was the other four problems (idle players, no
+  accumulating state, generic obstacles, lookup load). Always ask what the
+  request is *actually* for before designing to its literal wording.
+- **When a house subsystem gets a digital aid, test the aid headlessly.** A
+  ~60-line stub-DOM harness in Node drove the real console module through every
+  state transition (clamping, round cap, overtime drift, undo) with no browser
+  and no dependencies. Worth it: two of the "failures" it surfaced were the
+  harness lying, which is itself the thing you want to find before a
+  convention floor.
diff --git a/skills/rpg/references/playtest.md b/skills/rpg/references/playtest.md
new file mode 100644
index 0000000..bc6b5cb
--- /dev/null
+++ b/skills/rpg/references/playtest.md
@@ -0,0 +1,71 @@
+# Desk Playtest Protocol
+
+A scenario is not "convention-ready" until it has survived desk playtests and
+the fixes are applied and versioned. Three passes is the proven cadence:
+full session (GM chair) → stress re-run of the weakest act → full session
+(player chair, contrarian choices).
+
+## Method
+
+- **Full table simulation.** Run the scenario as GM *and* play every pregen as
+  a distinct player archetype. The proven quartet:
+  - the methodical veteran note-taker
+  - the roleplay-first talker who never reads handouts silently
+  - the quiet rules-literate tactician
+  - the lore-digger who asks "can I check…?" constantly
+- **Roll dice honestly** for every contested beat (physical dice or a fair
+  roller). Fudged playtests produce fudged timing.
+- **Log timing against the schedule** per section: planned vs actual, with the
+  drift explained. Note where the break landed.
+- **Play the players, not the script.** Ask the obvious professional questions
+  ("I scan her", "can I phone ahead?", "I run the face against city feeds").
+  Every improvised ruling the playtest forces is a hole in the text — the
+  written scenario must absorb it.
+- Later passes go **contrarian**: split the party wrong, skip the expected
+  route, refuse the set-piece, shoot early. Verify the fast-tracks and
+  contingencies actually work.
+
+## The playtest document
+
+One markdown file per pass in `playtest/`:
+
+1. **Method header** — what was simulated, which version of the scenario
+2. **Play log** by section — planned/actual minutes, what landed, what
+   dragged, table quotes worth keeping, each problem flagged inline with ⚠
+3. **Player-hat verdicts** — one line per player archetype, with a score
+4. **Feedback summary / fix list** — numbered, prioritised, each fix stating
+   the problem, the proposed text change, and where it goes. Nothing is
+   rewritten inside the playtest doc itself.
+
+## Applying fixes
+
+- Apply fixes as a separate pass; bump the version (v1 → v2 → v2.1 …) and
+  record in the README which playtest's fixes each version absorbed.
+- Classify every ⚠ before fixing:
+  - **Text hole** (a legitimate player move the scenario doesn't answer) →
+    write the answer into the scenario, not the GM's head
+  - **Timing** (a section runs hot) → pre-write the cut into the Pacing Note
+  - **Staging** (a scene lands only if performed a specific way — e.g. an NPC
+    must recite while the GM hands the card) → formalise the staging
+    instruction in the scene
+  - **Missing option** (players invented a compromise the text doesn't list) →
+    add it to the choices/outcomes tables
+- Timing target: full run inside the time budget with the drift noted. If a
+  playtest lands on schedule "to the minute with zero slack", that is a
+  **fail** for a real table of strangers — cut deeper.
+- Convention-ready = full-session playtests from both chairs inside budget,
+  multiple endings reached, art and handouts done, README status updated.
+
+## Post-slot debrief (after the real convention run)
+
+The con run is the highest-value data the system ever gets. Same day, ten
+minutes, one file in `playtest/` (`<scenario>-con-run-<event>.md`):
+
+- Timing actuals per act vs the table, and where the break landed
+- Every improvised ruling (each one is a text hole — classify as above)
+- Compromise endings or routes the table invented
+- What confused players, what they quoted back, which handout landed hardest
+- Verdict: what to fix before this scenario runs again
+
+Then close the loop as usual: fix list → version bump → transferable lessons
+appended to `lessons.md` → Graphiti.
diff --git a/skills/rpg/references/pregens.md b/skills/rpg/references/pregens.md
new file mode 100644
index 0000000..101598c
--- /dev/null
+++ b/skills/rpg/references/pregens.md
@@ -0,0 +1,47 @@
+# Pre-Generated Characters
+
+Pregens are half the scenario. A con player meets the game through the sheet.
+
+## The three tests (every roster must pass all three)
+
+1. **Rules-accurate.** Audit every sheet against the system's build rules —
+   attribute/skill budgets by experience tier, derived stats, restrictions
+   (e.g. Blade Runner replicant rookies carry no specialties). Show the audit
+   math in the pregens file so errors are checkable:
+   `(Budget: Veteran +2 attr → INT A, EMP B; +12 skill increases ✓)`.
+2. **A working party.** Each PC owns a distinct **investigation/action lane**
+   with minimal overlap (scene-reader, people-person, muscle/pursuit,
+   forensics/tech is a proven quartet). If the scenario has a set-piece, check
+   no PC sits skill-dead all night — rebuild the sheet or the set-piece.
+   Flag deliberate rebuilds of existing roster characters for review.
+3. **Wired into the case.** Every PC gets a **Private (do not share)** box: a
+   personal seam the scenario will pull on, phrased as a question the night
+   will force them to answer. The best private seams make the player *decide
+   something about themselves* in Act Three, not just recognise a connection.
+
+## Sheet contents (per PC)
+
+- Name, archetype, experience tier, one-line essence in italics — the essence
+  line doubles as the GM's spoken pitch during Slot Zero pregen picks
+- Attributes table, derived stats (show modifiers: *incl. replicant +2/−2*)
+- Skills (only the ones above baseline), specialties with rules text inline
+- Gear with the numbers the player needs (damage, range) — no page refs
+- **Signature item** with a small mechanical hook (e.g. interact once per
+  session to heal 1 stress) — these produce the best unprompted table moments
+- **Key memory** and **key relationship** — write relationships to NPCs and
+  places that appear in the scenario, so the case walks through their lives
+- **Private** box (see above)
+- **"Playing X"** — three bullet behaviour cues, imperative mood ("speak
+  second, watch first"), so a stranger can act the character in 30 seconds
+
+## Roster-level checks
+
+- Party seams: note which PCs pair naturally for splits, and give a
+  recommended 2+2 split per act (but let players choose at the table).
+- Player-count scaling: state which PC drops first at fewer players (their
+  lane must be recoverable elsewhere), how splits adjust (2+2 → 2+1), and
+  what a walk-up extra plays if the system supports a +1.
+- The roster file opens with goals + audit notes + errors fixed from previous
+  versions, so the next revision has provenance.
+- Produce print-ready sheets (HTML → A4, one page per PC + a table
+  quick-reference page). Portraits from the scenario art set.
diff --git a/skills/rpg/references/production.md b/skills/rpg/references/production.md
new file mode 100644
index 0000000..514f548
--- /dev/null
+++ b/skills/rpg/references/production.md
@@ -0,0 +1,71 @@
+# Production Pipeline
+
+Everything beyond the scenario text that makes a slot runnable.
+
+## Folder & repo layout
+
+Each scenario ships as a self-contained folder:
+
+```
+<SCENARIO>/
+├── README.md          # index table, status line (version, playtests, remaining prep), licence footer
+├── scenarios/         # the scenario markdown (+ art/ for scenario illustrations)
+├── characters/        # pregens.md, character-sheets.html, art/ (portraits)
+├── handouts/          # print-ready handout pack (HTML)
+├── reference/         # set-piece GM quick-reference sheets
+├── playtest/          # one file per desk playtest
+├── gm-utility/        # GM console app (module per scenario)
+└── print/             # final print artifacts
+```
+
+- iCloud is the working copy:
+  `~/Library/Mobile Documents/com~apple~CloudDocs/RPGS/Conventions/<Event Year>/<Day>/<Slot>/...`
+- Each event gets a **private GitHub repo** (`slaguru666/<Event><Year>`, e.g.
+  `Continuum2026`) as backup + cross-machine share. Commit at meaningful
+  checkpoints (draft, post-playtest, version bumps), not per-edit.
+- Licensed settings (Blade Runner, CoC, …) stay in **private** repos with the
+  licence footer: "© <publisher>. Private convention play materials — not for
+  sale or distribution."
+
+## Handouts & props
+
+- IDs: `[Slot]-H[act]-[letter]` (e.g. `F2-H00-A` = Friday slot 2, pre-game,
+  first handout). Every handout appears in the scenario's per-act handout
+  tables with When Given.
+- Build one **print-ready HTML pack** per scenario (A4, print margins,
+  background graphics on). Physical props beat screens: print guidance in the
+  scenario says what to hand around vs lay flat.
+- The best handout tricks are *staged*: duplicated documents revealed
+  simultaneously, sealed items ("bag it, don't open it"), cut-out props.
+  Write the staging into the scenario, not just the pack.
+
+## Artwork
+
+- Base style: **ink sketch / pencil illustration** — hand-drawn, single-weight
+  line, sparse wash, negative space. Never photorealistic. Each scenario adds
+  one art-direction line (e.g. "heavy rain hatching").
+- Asset IDs `[Slot]-ART-[##]`, tracked in the scenario's Artwork Assets table
+  with description, where used, and the full generation prompt (prompts are
+  part of the document — regeneration must be possible).
+- Generate with the Midjourney bridge: `mj-gen "<prompt>"` drops a PNG on
+  disk. Keep alternates in an archive folder; portraits feed the character
+  sheets.
+
+## GM console (`gm-utility/`)
+
+Touch-screen GM app (static HTML, no build step) with: session clock with the
+scenario's checkpoints, dice roller with system-specific push/advantage rules,
+set-piece runner (e.g. chase obstacles), PC/NPC dashboards, fullscreen props,
+autosaving state. Scenarios plug in as modules — when building a new scenario,
+add a module rather than a new app.
+
+## Print checklist (before the con)
+
+- Character sheets: A4 portrait, 100% scale, one page per PC + table
+  quick-reference page
+- Accessibility: handout body text ≥11pt (con halls are dim, eyes are tired);
+  the GM reads every handout aloud as it's given
+- Handout pack, including any duplicated staging cards
+- Set-piece GM quick-ref
+- GM read-through of the final versioned scenario
+- README status flipped to convention-ready with remaining-prep list emptied