I LOVE engaging with my Openclaw (formerly Clawdbot) agent, Frank, but I had him hooked up to Claude Opus and it was getting expensive! I used Frank as my golf caddy/coach the other day, and the post-round analysis report was awesome!!(future post) …but that entire session cost me ~$11.
So I switched Frank’s brain to kimi k2.5, a much cheaper LLM. I use kimi k2.5 through Fireworks AI, just for trust factor. So far, I don’t notice a difference in Frank, but my costs are way lower. But, I’m sure there are tasks where I may want to use a different LLM. For example, I might still want to use Opus for deep reasoning.
So I had Frank create a personas/ directory. Each .md file defines a character: their name, which LLM powers them, their personality, and what they’re good at.
Frank, not me, came up with these personas:
⦿ deep-researcher.md → Claude Opus, methodical, cites sources.
⦿ quick-helper.md → Kimi, fast, concise.
⦿ creative-writer.md (Scott)→ Opus, witty, stylistic.
Now instead of “use model X with temperature Y,” I just say: “Ask Scott to draft a post for me”. Scott actually did draft this, but I still do final edits myself.
The system reads the persona file, extracts the model config + personality, and spins up a sub-agent with that identity baked in. Like a film director casting roles — you don’t tell Meryl Streep how to act, you just cast her correctly.
The right AI for the right job.
My god, I’m going deeper and deeper into this AI rabbit hole and it is amazing. I’m telling y’all, I feel so much so much smarter with Frank and team by my side.
If you need help putting your team together, let me know, I’d be happy to help!
Engaged in a Perpetual Search for Knowledge
