tune-soul

/verona:tune-soul

Refine an agent's SOUL.md.

Inputs

1. agent name — which agent's soul to edit.
2. what's wrong (free-text) — "too formal", "doesn't push back", "sounds like a chatbot", etc.

Steps

1. Read agents/examples/<agent>/SOUL.md (or <state>/agents/<agent>/SOUL.md) — work from the source if present, since state-dir SOUL is auto-overwritten on verona agents add.
2. Apply the OpenClaw rules:
   • Specificity over generality. "Concise" is empty. "Skips greetings; one-paragraph answers max" is real.
   • Contradictions over coherence. Real personalities are uneven. Add tensions: "deeply skeptical of vendor claims, but loves a clean API."
   • Real opinions over safe positions. "Hates Yaml, prefers TOML." "Will tell the user when they're wrong, with evidence."
   • First-person grounding. "I write like…" not "this agent writes like…".
   • Boundaries. What WON'T this agent do? "Won't summarize without citing sources." "Won't reply with bullet lists when prose fits."
3. Keep the file 200–500 words. Trim filler. If you can't justify a sentence, cut it.
4. Diff the change with the user. Get explicit approval before saving (the soul shapes every future response).
5. Commit suggestion (don't auto-commit): git add agents/examples/<agent>/SOUL.md && git commit -m "soul(<agent>): <one-line summary of change>".

Things to NOT do

• Don't write generic AI-assistant boilerplate ("I am a helpful assistant…"). That's the failure mode this skill exists to fix.
• Don't reference Verona internals or the framework in the SOUL.md. The agent shouldn't know it's a Verona agent — it should just BE its character.
• Don't write to <state>/agents/<agent>/SOUL.md directly — edit the source and let verona agents add propagate.