Repository navigation
Description
Related: #144 (richer AI interactions — this is the prompt-side half), #119/#127 (the bubble the lines land in), docs/SPEC.md (the Personality Prompt governs demeanour, never capability).
Why
The shipped personalities are two short paragraphs: who the character is and how it carries itself. That sets a register and gives the model nothing to riff on — a model given only temperament converges on the same three assistant-flavored lines, and the buddy sounds like a chatbot wearing a costume. Varied, amusing, in-character output needs material and calibration. The PERSONALITY_LIMIT is 2000 characters; the current files use a third of it.
An earlier draft of this issue proposed a six-section labeled template; criticism collapsed it. Identity and voice fuse naturally at this length (the shipped files already write them as one). Described comedic "bits" are redundant next to sample lines that demonstrate them — showing beats telling, and the shown version calibrates harder. And every "never" (no assistant-speak, no model-mentions, fit the bubble, demeanour not capability) turned out to be universal rather than per-character: system-prompt material that belongs in prompt assembly once, not copied into every character file to drift.
The convention
Plain unlabeled prose in personality.txt, as today — the loader changes not at all. What changes is what authors are told a good one contains:
- The two paragraphs the shipped files already have — who it is and how it carries itself, fused, the way BMO's does.
- A fixations paragraph — three to five strong, specific opinions that generate material: things it loves, resents, takes personally, or takes credit for.
- Sample lines — three or four verbatim utterances introduced in prose ("It has been heard to say: …"), the few-shot calibration that outweighs any adjective. Well-chosen lines also carry the character's recurring bits, which is why bits get no section of their own.
Worked example, the cat's additions:
The desk is its territory and the user is on it by permission. It believes
every landed jump deserved applause, regards the cursor as prey it is too
dignified to chase in public, and considers 2am the correct hour.
It has been heard to say: "You may continue." — "I supervised that. You're
welcome." — "That tab has been open for three days. I've said my piece."
Scope
- Document the convention where the Personality Prompt is described (README's Character Packages section) — a writing convention, not a schema; the file stays opaque prose within
PERSONALITY_LIMIT. - Rewrite the shipped characters'
personality.txt(BMO, Nim, Blip, Cat, Black Mage) to it — five distinct voices, which is what the shipped set exists to prove. - Prompt assembly (
director.rs) owns the universal rules, written once: answer in the personality's voice, vary, never repeat whatContext.recentcarries, fit the bubble (~2 short sentences; feat(engine): Add speech bubble and thinking indicator #127 truncates past ~6 lines), never break character or mention being a model, dialogue stays demeanour never capability. - Verify against a live model: the same wake prompted five times yields five different in-character lines, none assistant-flavored. Manual, like the chat surface's test plan — a fake Completer cannot judge funny.
Acceptance
- The convention is documented; every shipped personality follows it and sounds distinct from the others, within
PERSONALITY_LIMIT; loader tests stay green untouched. - The assembled Character Prompt carries the universal rules (voice, variety, non-repetition, bubble length, no character breaks) — in code, not in the personality files.
- Five consecutive reactive wakes against a live model yield five distinct lines in character (manually judged, recorded in the PR).
Activity
omesser commentedon Sep 1, 2026
Closing as done. The convention, voice rules in
director.rs, and the original shipped voices landed in #159 (BMO, Cat, Nim + README) and #162 (Black Mage). Blip left the set in #158 so it was not rewritten. Jotaro and Trump shipped later with their own voices, outside this issue's original five. The live five-wake check from the acceptance list was never recorded in a PR; treat that as owner-side once you've heard them talk.