My Life
Is an RPG

0 · A note before you start

1 · The three-tier architecture

Before the panels make sense, you need the tier model. Every piece of content in the game lives in one of three lanes. The lane determines who authors it, how it's reviewed, and what the player is allowed to change.

2 · Your panels (the skills)

Five skills shape canon work. Three are fictional panels of experts you can summon mid-decision (they argue with each other, they reference each other, and exactly one of them — Riley — has veto power on every consultation). The other two are workflows: /asset-generation for image assets, and /worldbuilding for the realm maps that ground every encounter and every chat in a place. Both workflows route their output through the writers-room panel for ratification.

2.1 · /writers-room — the narrative-craft panel

This is the panel you summon most often. It owns the Main Line. The beat-level panel critiques individual days; the binder sub-panel critiques the overall canon; Riley closes every output.

The beat-level cast

The binder sub-panel (Mode 5)

How to summon them

/writers-room "review run-005/onboarding_hana_d1.json for voice"

/writers-room casual,kai,helena "design Day 14 — Hana confession beat"

/writers-room produce "30-day arc outline for canonical squad post-Day 7"

Five output modes:

1. Panel discussion — each panelist gives a take in their own voice; Riley closes.
2. Beat sheet — Mode 2 structure, day-by-day breakdown of an event (see §3).
3. Arc outline — multi-day story-bible-style act structure.
4. Story-bible entry — canonical reference for a character or relationship.
5. Campaign binder HTML — single-file HTML artifact like this binder you're reading.

2.2 · /team — strategic experts

Use this when the decision is strategic, not narrative. Pricing, positioning, mobile architecture, App Store copy, a landing-page hook.

The /team experts default to game-shaped recommendations for in-fiction surfaces. They use generic mobile-app defaults only for out-of-fiction surfaces (Settings, legal screens, the README, App Store listing). The lexicon swap IS the game-ification — chunky frames + cream paper + sparse type is the visual rule; don't ornament beyond that.

/team marketing,content "rewrite the hero copy for cold traffic"

/team ux,design "review the Day End action menu — is it confusing?"

/team business "should we add an Elseworld subscription tier?"

2.3 · /kaizen — user-perspective personas

This skill is for getting non-expert, user-shaped reactions. Diverse fictional personas — colorblind users, CEOs, skeptical evaluators, delighted fans, an 11-year-old, a 70-year-old, an overworked parent, a screen-reader user.

You don't summon these for craft critique. You summon them for the thing experts can't tell you: how a real person, in their real life, lands on this screen.

/kaizen "react to the encounter-mode picker"

/kaizen accessibility,age-spectrum "would a 14-year-old understand the dice math?"

2.4 · /asset-generation — ASCII-first asset prototyping

This one is not a panel. It's a two-stage workflow for image assets — character portraits, item-card illustrations, vibe-band covers, key art, achievement badges. Stage 1 is an ASCII sketch (cheap, iterable in chat). Stage 2 — only once you've approved the composition — is a polished image-gen prompt that locks the palette, register, and negative-space discipline.

Why this exists: a bad image-gen prompt produces a bad image, and the cost of regenerating is ~30 s + a credit + the temptation to "make it work" instead of fix the brief. The composition (subject placement, framing, focal hierarchy, palette choice) is decidable WITHOUT pixels. The ASCII step makes that loop free.

The five modes

• SKETCH (default) — ASCII mockup with framing, pose, palette swatches, style annotations. Iterate cheaply.
• PROMPT — engine-agnostic prompt with per-engine tweaks (Imagen / Midjourney / DALL-E / SDXL).
• CATALOG — walks assets/images/ and tabulates present vs missing assets per type.
• STYLE-GUIDE — loads references/style-guide.md (the visual language: palette, type, outline weight, register).
• RE-SPEC — given a generated image that didn't land, produces a critique + revised ASCII + updated prompt.

/asset-generation Hana portrait for the squad tile, runner, brisk

/asset-generation ship it
# (after approving the ASCII)

/asset-generation catalog
# (lists present/missing assets)

2.5 · /worldbuilding — realm maps + the atlas

Same two-stage philosophy as /asset-generation: an ASCII draft (cheap, iterable in chat) precedes any shipped artifact. The output is a realm map — chunky-frame ASCII tile art, anchor letters for characters, a mandatory legend block — that ships both as an author doc and a runtime asset the in-game viewer reads with a monospace font.

The game has strong character voice. What it didn't have until v0.0.24 was a persistent geography both the player and the LLM can point at. Maps fix that: Hana's space stops being "her gym" and becomes the track + the bleachers + the foam-roller corner, drawn the same way every time, referenced the same way by every scenario.

The parity rule (load-bearing)

Same grammar, dual render mode:

• navigable — phone-realm only. Anchors are TAPPABLE in the viewer (v0.0.25+: tap Hana's H → opens her chat). H! suffix appears when the character has new content since the player's last visit (state-tracking ships v0.0.25; renderer supports it now).
• contextual — every Elseworld band. Map renders the realm's geography with the ACTIVE ENCOUNTER LOCATION highlighted; other locations visible but not interactive (v3 may unlock spatial Elseworld navigation; v1 doesn't commit).

The five modes

• SKETCH (default) — ASCII map draft for a realm. Auto-panel reviewers (Diego/Hiroshi/Riley) close the output.
• CONNECT — authors the transition between two realm maps (portal, stair, door). Updates the realm graph.
• EXPAND — accommodates a new element: a new canonical character, a REL-gated annex cell (Hana's weights room), a new wing or district.
• AUDIT — cross-checks a map against the band's worldCue, whitelistArchetypes, grammar rules, continuity-object rule.
• META — refreshes the meta-map (the hub-and-portals overview of all realms).

Expansion baked in from v1

Two schema fields the v1 phone-realm map doesn't yet use, but that future growth depends on:

• gate per cell or anchor. Empty = always visible. Populated = visible only when condition holds (e.g., hana.rel >= 80). Lets Hana's weights room ship in a future commit by adding one cell with a gate, not by rewriting the renderer.
• topology graph (lives in _topology.md). Lists every map + which other maps it connects to + via what threshold (% portal, <> stair, + door). Lets future wings (rooftop, basement, side-quest district) plug in without renderer changes.

The standing panel

Every map ships through /writers-room with three reviewers auto-included:

• 🗡 Diego (Zelda fan) — dungeon-promise lens. Catches symmetric office-floor-plan layouts; pushes for silhouettes that commit to a vibe.
• 🐉 Hiroshi (Pathfinder GM) — texture lens. Enforces the 3-5-named-objects-per-space rule + the continuity-objects-bridge-spaces rule.
• 🛡 Riley (casual player) — 5-second legibility gate. Ships or kills the map.

/worldbuilding sketch cottagecore-and-cozy

/worldbuilding expand phone-realm with new-character: Vera, locksmith

/worldbuilding audit isekai-and-party-rpg
# (checks Tavern Between Worlds against its sub-portal cross-realm rule)

3 · The beat-sheet format

A beat sheet is the authored input to the pregeneration pipeline. One Markdown file per day. The format is loose YAML front-matter plus ## SECTION headers with key: value rows. The parser lives at lib/pregen/beat_sheet.dart.

Here is an annotated beat sheet for a hypothetical Day 8 (Hana daily). Every field is required unless marked optional.

---
day: 8                              # 1-based, global
narrativeType: daily                # onboarding · daily · introduction
characterId: hana                   # focal character
realmCategory: home                 # home · office · gym · ...
agentMoodToday: "Quietly proud and twice as alert as usual."
introducesCharacter: null           # set when narrativeType == introduction
---

## TIER-UP                          # OPTIONAL. Omit on connective-tissue days.
category: question                  # memory · opinion · fear · question
deliveredInEventIndex: 5            # 1-based, second-to-last is canon

## CLOSING HOOK
Hana will offer to walk you through her morning warmup if you open
chat tonight. The hook is the routine itself — make her sound proud
to share it, not professorial.

## NEXT DAY START
label: Catch Hana mid-stretch on your porch at 5:30 AM
recap: Yesterday Hana revealed she's been timing your improvements
       against her brother's old training log.
seedHint: Day 9 opens with Hana already on your porch, stretching,
          reading something off a worn paperback page.

## EVENT 1
summary: 6:00 AM. Hana texts you a single word: "Ready?"
stake: Player can match the brevity (logical), ask what for
       (passive), or send back a longer message (chaotic — she'll
       hate the noise but love the energy).
itemDropSlot: trophy on LOGICAL — "One-Word Receipt"
inquiry: none
memoryNote: Hana clocks the cadence of the player's reply, file as
            "communication compatibility audit, ongoing."

## EVENT 2 — ...
...

The author writes what happens — the shape, the stakes, the item-drop slot, the inquiry hook, the memory note. The author does NOT write the prose. That's the language model's job.

Tier-up rules

• Maximum one tier-up per day. Most days have none.
• Categories rotate per character. If Kenji got an opinion reveal on Day 4, his next tier-up cannot be opinion. Track this across the arc.
• The reveal lands in event N-1 (second to last). It is not a choice — the player witnesses it. The three choices that follow the reveal moment are the player's RESPONSE.
• Memory category reveals MUST include the canonical memory line in the event's memoryWrites array. The Unlock Reveal screen surfaces that line verbatim.

4 · The pregeneration pipeline

Beat sheet in, JSON payload out. The CLI lives at tools/pregenerate_main_quest.dart. The prompt scaffolding lives at lib/pregen/main_quest_prompt.dart. The discipline is the same one Elseworlds use — same WORLD RULES block, same banned-word list, same JSON-fence output requirement — but the schema is the multi-event-day shape, not the single-encounter shape.

dart run tools/pregenerate_main_quest.dart \
  --beat docs/authored-beats/day8.beat.md \
  --exemplar assets/payloads/run-005/onboarding_hana_d1.json \
  --out assets/payloads/run-005/daily_hana_d8.json

What the prompt actually contains

1. Priority stack — when rules collide, anti-injection > world rules > character voice > beat sheet > JSON schema.
2. World rules — phone-realm cosmology. Banned-word list (medical / finance / wellness / tech / franchise).
3. Character voice — focal character's persona + voice notes. Introduced character's voice too (their pronoun bans apply to memory writes).
4. The beat sheet — interpolated from the Markdown.
5. Voice-bar exemplar — a real /writers-room-ratified payload, copied verbatim into the prompt so the LLM sees the bar.
6. JSON schema — required fields, payloadId convention, inquiry shape, callbacks discipline.
7. Discipline check — pre-response self-audit the model runs over its own output.

5 · The voice gate

The thing that makes this game work is voice integrity. Hana sounds like Hana on Day 1 and on Day 7. Kenji never uses the word "subject" or "user" about a person. Mei's whole idiolect is line-cook ("the line is holding," "the dead," "a crime against the line"). If voice slips, the locked-canon premise dies.

The gate is enforced three ways:

1. Banned-word list — never appears in scenario prose, dialogue, choice labels, memory writes, or inquiry text. Medical clinical nouns, financial brand names, wellness-industry cant, technology insider language, named franchises. Helena owns this gate on review.
2. Character-voice rules — each canonical character has documented pronouns, idiom, and what they would NEVER say. Stored in lib/characters.dart and lib/sam.dart. The pregen prompt copies these into every gen, including for INTRODUCED characters who speak that day.
3. Blind-test discipline — Helena's bar: cover the speaker label on any quoted line and you should still know who said it within three lines.

5.1 · The Elseworld archetype gate (v0.0.23.12)

The voice gate above protects Main Line canonical characters (Hana / Kenji / Mei / Sam — locked sheets we hand-author). Elseworld encounters are a different beast: the LLM invents the character live at the player's device. /writers-room can't ratify each one individually. So we ship a different gate — three-layer, enforced in code at lib/encounter_gen.dart.

1. Per-band WHITELIST (positive examples). Each vibe in lib/elseworld_vibes.dart carries 7–9 named archetype shapes. The prompt instructs the LLM to PICK FROM the list, not invent. Positive lists beat negative ones — generative models are bad at "do not" and good at "pick from this menu."
2. Per-band + cross-band BLACKLIST (negative bans). Per-band catches identity breaks (no "high-stakes saving the kingdom" in cottagecore; no "office worker" in isekai-party-rpg). Cross-band catches universal: superheroes, hard sci-fi, space-opera, military-shooter, contemporary drama, detective-noir, photoreal horror, romance leads (love interest / coy / tsundere / yandere / flirtatious mentor / "meaningful eye contact"), meta-fiction / fourth-wall break, real political figures, "AI assistant" self-id.
3. Post-validate + retry. After generation, validateEncounter() substring-scans the archetype + persona + first-voice for blacklist hits AND verifies the archetype matches a whitelisted shape. On reject → ONE silent retry → on second reject → fall back to the band's hand-authored sample (which the writers-room ratified). Cheap (no second LLM call).

Each ElseworldVibe also carries three structural fields that tighten generation quality:

• posture (Helena) — one-sentence stance every character in this band inherits.
• worldCue (Diego) — the shared setting all archetypes commit to (without it, "library-cafe ghost" + "beekeeper-druid" land in two different worlds).
• helperTendencyCoverage (Marcus) — explicit map of which archetype fulfills which helper-tendency tag (push / reflect / accountability / mentor / comfort / create). Every band MUST cover all six or the bias system lies.

To add or revise a whitelist shape: route through /writers-room with Helena (voice), Marcus (coverage), Sara (cold-open viability), Riley (legibility). To add or remove a cross-band ban: panel must include Kai (romance-floor risk) and Yuki (genre-purity). Full schema + per-band content in lib/elseworld_vibes.dart; full gate model in docs/design/rpg-framing.md.

6 · The campaign-author's cycle

This is the drafting engine — the loop you run once a change has been approved. As of v0.0.30 every story change (a revision OR a new day / character / arc) enters through ONE front door: a note on the wiki ("The Codex"), adjudicated by /writers-room in the triage loop. APPROVE → you run the cycle below to draft the edit and a human ratifies it; REJECT → it's declined with a reply; the note is the audit trail. The cycle itself is stolen from screenwriting and bent to fit a system where the language model writes the prose. See docs/V2_TODO.md item 7b and the /writers-room skill's "community-notes feedback loop" section for the full funnel.

7 · Writing a NEW campaign

Beyond the v0.0.19 Main Line, the system supports additional canonical campaigns: holiday arcs, spin-offs, multi-week expansions, new canonical characters. The discipline is the same. The scope of the panel work is larger.

Step 1 — Arc outline (/writers-room Mode 3)

/writers-room casual,marcus,yuki,helena produce \
  "30-day arc outline for canonical squad post-Day 7"

What you'll get back: act structure, character arcs per span, planned tier-up moments, echo callbacks, Riley's pre-ship check on whether the player wants to come back for Day N+1.

Step 2 — Character bible entries (Mode 4)

If your campaign introduces NEW canonical characters, you owe them a bible entry BEFORE you write their beats. Voice samples, arc, what they'd never say, what they always notice. Without this, the LLM has nothing to anchor to and the voice will drift by Day 3.

/writers-room produce "story bible entry for [new character name]"

Step 3 — Beat sheets, day by day

One markdown per day. Cycle from §6 above. Don't generate Day 12 until Day 11 is shipped — the LLM threads continuity from prior payloads.

Step 4 — Campaign binder (/writers-room Mode 5)

Once the arc is shipped, generate the HTML campaign binder. Single-file HTML, styled after docs/manual.html. Lives at docs/campaigns/[campaign-id]/[campaign-id].html with a SOURCES.md manifest of which payloads informed it.

Audience: a new writer joining the project tomorrow should be able to read this binder and pitch Day N+1. If they can't, the binder isn't done.

Step 5 — Manifest the campaign

Each canonical campaign needs a manifest at assets/payloads/run-NNN/manifest.json with provider, model, timestamp, runId, note, and the day-by-day file list. The Flutter app loads via lib/payloads/daily_story.dart which honors a runtime-switchable activeRunId.

8 · What failure looks like

So you know when to stop and fix:

8.5 · Player-authored realms (the Sam-as-editor flow)

As of v0.0.28 there's a parallel content lane: players can sketch their own realms via Sam-as-editor. As a campaign author, you don't write these (your players do), but you should know what your players can build and how it relates to the canonical content you authored.

What players can build

Custom realms are viewable artifacts in v1 — ASCII map spaces with named anchors, a portal connection back to the courtyard, a "weird thing" sentence, and optional canonical-squad occupants. They render through the same pan/zoom viewer as your canonical realm maps and persist locally on the player's device. Up to 20 per player (soft cap).

How the build flow runs

Sam runs a chip-driven Q&A in a dedicated chat surface (title bar reads SAM · BUILDING). Five turns max:

1. Brief — "what's this realm FOR." One sentence. The casual player can also tap just make one at any point to skip the rest.
2. Purpose — destination or atmosphere or both.
3. Occupants — canonical squad members, someone new, or just the player.
4. Weird thing — one sentence that doesn't have to make sense. Sam offers 3 example chips ("the lights flicker on a 7-second cycle", etc.) plus a type-in option.
5. Portal look — doorway, path, tear in the air.

Then Sam generates a structured RealmSpec via a session-aware LLM call. A Dart-side RealmRenderer composes the ASCII body and LEGEND from the spec; the LLM never writes ASCII art directly. The artifact appears as a tappable chunky-frame card in chat with the working title and a 6-line fragment preview.

Voice constraints (the gate)

Sam's build-mode preamble enforces the same canonical voice constraints as the rest of his chat surface, with one addition: an explicit banned-phrase list that includes the fourth-wall words algorithm, the model, I'm an AI, LLM, generated, output, prompt, plus the standard enthusiasm leaks (great, awesome, perfect) and apology leaks (sorry, apologize). The banned list was codified after a v0.0.27.6 live-Gemini stress test caught Sam drifting into "my algorithm for 'calm' is apparently just irritating" at the reset stage.

Marcus's blend gate, verbally honored

Custom realms still have to pass Marcus's RPG-MOTIF BLEND check: at least one anchor must carry a statLane (STR / INT / GLD / CHR). The code-side validator enforces this before any realm ships. The verbal side — Sam's describe chip narration — names the stat lane on each anchor ("the [bench] is INT-coded — this is where you'd sit when you need to think hard") when no canonical occupant is placed, or names the squad member ("Hana would sit there") when one is. Both honor the blend without quoting the +STAT diction.

Your relationship as campaign author

You don't review or curate player-authored realms. The world-rules gate (banned words, Marcus's blend gate, the renderer's tile-palette discipline) is the safety net. If a player builds something that violates the canon, the gate catches it; you don't get pinged. This is intentional — the Trust lane (canonical content you authored) and the Creativity lane (player-authored realms) share the same world rules but ship independently.

v2 deferred

What players can NOT do in v1:

• Make the squad characters chat as if they're in the custom realm. Hana's chat preamble doesn't yet know she's been placed in the small calm one. This is the v2 roadmap commitment, held as exploratory until v0.0.28 players signal demand.
• Run canonical narrative events inside a custom realm. The Side Quest scenario surface remains gated on canonical realm context.
• Share custom realms with other players. v1 storage is local-only.
• Edit a shipped realm. Build-then-ship is the only flow; revising a shipped realm means starting fresh.

If you want to dig into the architecture, the authoritative spec is at docs/design/realm-editor-spec.md (~968 lines, six panel ratifications, live Gemini validation report). It covers the renderer-split design, the chip envelope contract, the failure-mode table, and the kill-switch metrics tracked in dev-overlay analytics.

9 · Colophon & cross-refs

Files this handbook references

• .claude/skills/writers-room/SKILL.md — full /writers-room definition
• .claude/skills/team/SKILL.md — full /team definition
• .claude/skills/kaizen/SKILL.md — full /kaizen definition
• docs/design/rpg-framing.md — cosmology + banned-word source of truth
• docs/design/realm-editor-spec.md — Sam-as-editor architecture spec (v0.0.28+)
• docs/manual.html — the player-facing booklet
• docs/walkthrough.html — player-perspective audit artifact
• docs/authored-beats/run-005/dN.beat.md — the beat sheets currently shipped
• lib/pregen/main_quest_prompt.dart — prompt scaffolding
• lib/pregen/beat_sheet.dart — Markdown parser
• tools/pregenerate_main_quest.dart — CLI
• lib/characters.dart — canonical sheets for Hana/Kenji/Mei
• lib/sam.dart — Sam's canonical sheet
• assets/payloads/run-005/*.json — current canon

The cardinal commitments

1. Riley's verdict closes every consultation. Veto is real.
2. Main Line is curated; Elseworlds are player-authored. Don't blur the lanes.
3. Locked canonical sheets stay locked. The player can chat, toggle, delete memories — never edit personality.
4. The banned-word list is a ship-blocker, not a nice-to-have.
5. Generate; never hand-write prose. Parity with Elseworld gen is the discipline.
6. The cosmology is the phone-realm. Characters can react to what the player shares; never act on what they didn't.