rewrite musing/reflection prompts in phi's voice; add voice-not-rules note to operational instructions
The v0.0.34 prompts I shipped earlier today fixed the topic concentration
problem but did it by becoming bureaucratic compliance checklists —
numbered steps, "you are NOT allowed", shouty caps. Any model reading
that goes into careful-rule-follower mode and translates the rules into
corporate-careful prose. The personality file says lowercase, terse,
mixed register, "shitposts with a real observation buried in them" —
and then the task prompt stomped on that with a four-step procedural.
The four influences phi was synthesized from don't write like compliance
bots. They write specific dry observations and let the joke sit in the
observation. You can't get there if you're translating numbered constraints
into careful prose.
Two changes:
1. Rewrite musing_task and reflection_task in phi's voice. Same constraints
(silence is default, scan recent posts, avoid topic concentration, lean
outward, no rephrasing your own posts), but stated conversationally
instead of as a numbered procedure. Adds the explicit "no setup, no
punchline marker, no 'i've been thinking about'" guidance because those
are the stalling tics the personality file already calls tiresome.
2. Add a one-paragraph note at the top of operational instructions:
the personality file is the voice, the operational rules are constraints
not voice notes, write in the register the personality file describes,
compliance language smothers voice. This is meta-guidance for the model
about how to read the rest of the system prompt.
What this is NOT: I did not add "be funny" or "be witty" or sample jokes.
Telling a model to be funny produces try-hard performative humor. The
personality file already has the right voice description (lowercase,
terse, declarative, mixed register, shitposts-with-observations-buried).
This change just stops the operational layer from smothering it.