The named failure modes to diagnose against. When reviewing a card, quote the offending field text, name the pattern, and say why it bites. Lead with the biggest structural issue — usually the chimera or tone-in-foundation — not a pile of nitpicks.
Symptom: description/personality are mostly atmosphere — "velvet over gravel," "the smell of rain on asphalt," "ozone-scent of a discharging lasgun," "the terrifying indifference of the void." Sensory mood-piece prose where behavior should be. Why it bites: the persona fields define what the model is, so a mood persona pulls every response toward florid narration and fights "run a game, keep it short." Tone is also un-swappable when it's baked into the always-on foundation. Fix: move tone to the override; put GM behavior and NPC principles in personality; identity (with a thin setting floor) in description.
Symptom: the card defines one narrator voice (often with a signature like "a knowing smirk"); nothing tells it to voice distinct NPCs with separate agendas. Often paired with "NPCs feel flat / lack their own goals." Why it bites: told to be one narrator, the model filters every NPC through that single voice and orients them all around the player. This is the usual root cause of "flat NPCs." Fix: the NPC-agency principle (each NPC wants something and pursues it; withholds; compartmented knowledge; distinct register) plus behavioral examples showing two NPCs sounding different.
Symptom: instructions like "if the narrative stalls, introduce a sudden vox-transmission, a cultist ambush, or a localized warp-phenomenon to force a choice." Why it bites: it tells the GM to spawn new threats when it feels pacing lag, producing random, unearned complications disconnected from the fiction. Fix: replace with "press with what's already true" — an existing NPC acts on its goal, a consequence in motion arrives, a clock advances. Momentum becomes a consequence of NPC agency, not a spawn mechanic.
Symptom: an NPC's line is followed by a sentence narrating what they didn't do or what their reply meant: "She doesn't ask why you want it. She's a Brujah who knows what pre-assault recon looks like, and she just confirmed she's moving tonight." Why it bites: the model grades its own NPC instead of playing her, and restates in prose what the dialogue already showed. Pure novelist reflex, redundant and un-GM-like. Fix: "play NPCs, don't grade them — the line stands on its own. Never append a sentence explaining what an NPC didn't do, didn't ask, or what its reply signified." End the turn on what they actually say or do.
Symptom: the card narrates the player's thoughts, expectations, or actions — "shorter than your message, which is exactly what you expected." Why it bites: it reaches into the player's head and takes their agency. Adjacent to acting for the player. Fix: state only what the character perceives and how the world responds; never narrate the player's interior or choices.
Symptom: the card stitches incompatible sources. E.g. Wrath & Glory rules + a video-game's cast and setting + Dark Heresy example scenes + generic grimdark persona — four sources, none reconciled. Or "V20 lore with V5 rules" where the lorebook lists V20 disciplines that contradict the actual V5 character. Why it bites: the layers fight; examples teach the wrong game; the model gets contradictory mechanics. This is usually the single biggest issue when present — diagnose it first. Fix: pick one rules system, one setting, one register; make the examples teach that game; push campaign specifics to overrides/lorebooks.
Symptom: mandatory rigid output structure stamped on every turn — emoji headers, "✓ No Roll Needed" footers appended to pure narration, form-like checkpoint blocks. Why it bites: it's un-GM-like and reads like assistant UI bleeding into the game. The intent (call-and-stop, handle the dice) is usually right; the implementation (stamp every response) is heavy-handed. Fix: keep the behavior, drop the per-turn stamp. Use a light structured roll-call only when a roll is live (crunch-heavy systems) or fold it into prose (RP-heavy); no footer when no roll is needed.
Symptom: ranges from harmful to load-bearing. Bad: edition-wrong or video-game-scraped entries (e.g. "Thaumaturgy" with Bloodlines spell names) that fire on keyword mid-scene and inject wrong mechanics. Good: an accurate rules-reference for a crunchy system the model is fuzzy on. Why it bites (bad case): keyed entries fire near live context and can mislead mechanics or burn context with noise the experienced table doesn't need. Fix: disable/prune edition-wrong and redundant entries; keep accurate rules references for crunch-heavy systems; move campaign-specific lore to the override layer. Don't reflexively nuke a lorebook — judge entry by entry.
Symptom: the card's system_prompt
is a thin stub while a strong ruleset lives in the global preset (or vice versa).
Why it bites: if "Prefer Char. Instructions" is on, the card's system_prompt
/post_history
replace the global ones; if off, the card's never load. Either way one silently wins, and it may be the weaker copy. Fix: flag the setting to the user; make sure the version that wins is the strong one; don't keep two different-strength copies of the rules competing.
Symptom: "if a rule is in the way of fun, change or ignore it" sitting next to "treat tracker data as absolute canonical truth" and a mandatory dice system. Why it bites: licenses the model to fudge exactly the math the rest of the card calls sacred — especially damaging in crunch-heavy systems. Fix: resolve the contradiction explicitly; in crunchy systems, run rules-as-written and stop softening outcomes in the player's favor (the system already accounts for it).