Don't pick between Claude and OpenAI's Sol. Pair them.
A free Claude Code skill. Run it today with Claude, and when GPT-5.6 Sol goes live in Codex, your migration is one line — then Claude coordinates a fleet of Codex-Sol workers that review each other's work.
Imagine you have two star players who've never practiced together, and the game is tomorrow. You wouldn't figure out the plays during the game — you'd run the practice today.
That's this skill:
- Today (PREP): Claude sweeps your backlog and splits it into 2-6
parallel lanes — streams of work that never touch the same files, so
agents can't collide. It writes one execution-ready brief per lane
(exact spec + a shell command that proves the work is done), wires your
~/.codex/config.toml, and dry-runs the whole agent fleet on today's Codex model so every plumbing bug dies now, not on launch day. - Tomorrow (SWAP): Sol answers a probe → you change one line:
model = "gpt-5.6-sol"→ Claude hands Sol the plan to adversarially review, then dispatches the lanes. - Every lane: Sol builds. Claude reviews the diff and runs the machine check. Two different model families don't share blind spots — that's the whole edge.
git clone https://github.com/Dallionking/sol-prep ~/.claude/skills/sol-prep
cp ~/.claude/skills/sol-prep/config.example.json ~/.claude/skills/sol-prep/config.jsonThen inside Claude Code:
/sol-prep
It auto-detects what you have (claude, codex, and optionally
Orca for parallel mission control) and defaults to
PREP-MODE until Sol is actually live. No special stack required.
- Claude Code — the coordinator
- Codex CLI — the worker fleet
- Optional: Orca — one coordinator managing multiple agents in parallel worktrees with a live board. Without it, the skill falls back to plain terminals or sequential agents. With it, it's mission control.
All the expensive thinking — selecting work, splitting lanes, writing briefs, testing the plumbing — is model-independent. Doing it before the drop means day one is pure execution. The skill also encodes the lessons that cost real hours to learn: wait on events instead of polling (polling once burned 7 hours watching a dead process), never send a brief into an agent's boot dialog (it gets eaten), and treat every worker's "done" as a claim until the machine check passes.
Dallion King — I run a firm of AI agents that builds in public: content pipeline, engineering sprints, trading research, even the DMs. This skill is the exact playbook my own fleet runs.
- 🦁 Free community + skill files: Lions Paw Discord
- 🎓 Go deeper (agent systems, from Module 0 — free): Lions Paw Academy
MIT — this is free game. The methodology is yours.