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.

2027-01 prep (2026-07) — continuity searches for returning NPCs

Grep the role, not just the name. A returning NPC brief said "Witchfinder Garrett". Case-insensitive greps for garrett and garret both found nothing across every convention folder — the character is Garratt, from The Princes Bride. One wrong consonant nearly produced a scenario that contradicted its own prequel. Search the role (witch finder, witchfinder, inquisitor) and read the prior scenario's NPC roster table before concluding a character is new. Report "not found" as provisional, never as fact.
The prequel's early-finish Tangent is the sequel's premise, already written. Every house scenario carries a bonus scene for tables that crack it early. Those scenes are hooks with the cost already priced — The Princes Bride's crossroads offer (a ledger's blank column, a list of three villages, and "teach me what the meadowsweet meant") gave the sequel its inciting incident, its antagonist's arc, and its hardest ending, for free. Check the Tangent first when writing any follow-up.
A sequel that reuses the prequel's pregens should be built on their documented seams, not on the plot. The prequel's Sanction-rhyme table already paired the forger with the evidence-making witch finder; building the new case so that pairing is its best scene costs nothing and is the whole payoff for returning players. Keep the house rule intact — every seam still needs a written alternate route so new players lose nothing but the echo.

2026-08 — OPEN DAY (The Custodians prologue, house system)

Write the recruitment prologue out of the pregens' own biography lines. The ten "(before)" sheets each ended with a How they found her/him line. Two of them — a building that "stopped being on the street it was on" and a caller talked down "for eleven minutes" — were already the same incident, and the other eight fitted over it without a single edit to the plot. Build the prologue at the intersection of the existing seams and the scenario cannot contradict a sheet, because the sheets came first. Read the whole roster's private lines before inventing an inciting incident.
Reconcile the seams in both directions, and prefer changing the sheet. One line said night and the scenario is an afternoon: one word on the pregen was the cheap fix, and both became true. Where a seam needed an event the scenario did not have (a casualty for the one who carries people out, an injury for the doctor), the honest move is to add the event, not to bend the seam.
A scenario where nothing is hostile needs that stated as a GM instruction, in the first page, three times. The default GM reflex under pressure is to make something menacing. "Every figure is busy, polite or preoccupied" is a rule, not flavour.
Let them succeed and let the outcome not change. For a fixed-timetable act, write explicitly that every intervention works mechanically and alters nothing — and that the GM must not have them fail instead. Failing is comforting, because failing implies succeeding would have helped.
Three unresolved things is the right number. One is a loose end, two is a mystery the table will try to solve, four is untidy. Name them in the GM notes as do not resolve so a helpful GM does not answer them.
Midjourney will not stop putting lettering on a building. "no text / no signage / no lettering" failed twice on a shopfront-shaped prompt. What worked: describe it as a two-ink screen print, name the two inks, say "flat pale yellow background instead of sky", and add "no shops". Palette drift and hallucinated signage are the same failure — the prompt was letting the model reach for a naturalistic reference photo.

2026-08 — LAST ADMISSION (simplifying a prologue that was too complex)

"Too complex" usually means too many things the GM must hold, not too much text. OPEN DAY's load was a minute-by-minute timetable to narrate, three deliberate non-answers, and ten serial one-on-one scenes. The rewrite kept the same premise engine and cut to: one building, one truth, one decision, and three antagonists who each do one printed paragraph. Count the things the GM must track, not the pages.
Give an antagonist a page and let the players read it. Printing the monsters' backstory as an in-world poster — and telling the GM to say it out loud — turns a puzzle into a game about scripts. Each one then gets a "script break" (put the person back in the room; say his name; go indoors) that is a complete answer costing nothing. Players who find one feel clever; the fight stays available for players who want it.
A haunted attraction is the cheapest way to hide the real thing. A building full of paid performers pretending to be ghosts means "which of these is an actor" is the entire investigation and needs no explaining. It also supplies bystanders, a floor manager, a body, and a reason the players are all present.
Midjourney cannot count. Two attempts at "exactly four name plates" both returned three, and the second also drifted photoreal. If a count is load-bearing in the fiction, it belongs in a handout, never in generated art — a picture that contradicts the clue is worse than no picture. Cut it and reassign the spare plate.

15 KiB Raw Blame History