I stopped looking for a single perfect AI coder and combined a web UI and CLI (xor01.substack.com)

The author abandoned the search for a single “perfect” AI coder and built a practical, mostly-free “dual‑brain” workflow: use a powerful web UI as a Deep Thinker to ingest full repo context and produce a detailed, model‑agnostic implementation plan, then hand that plan to a fast CLI agent as the Fast Hands to execute changes reliably. This solves common failure modes — slow web CLIs, limited free tiers, and brittle autonomous agents — by separating heavy contextual reasoning (Gemini Pro/GLM in AI Studio, Perplexity for research) from low-latency, deterministic edits (Cline CLI with the free Code‑Super‑Nova model). Technically, the stack bundles a repo with repomix (repomix --output all_code.txt), uploads that bundle to a web chat to generate plan.md, then runs cat plan.md | cline -y --no-interactive to apply changes. The author tunes Cline with higher thinking budgets (cline c set plan-mode-thinking-budget-tokens=10000 and act-mode-thinking-budget-tokens=10000) and authenticates via cline auth to access free models. In one real example, Gemini Pro produced a 400‑line migration plan in ~3 minutes for a 3k‑line project; Cline executed it in ~5 minutes, creating DB tables, modifying endpoints and adding tests. The approach highlights a pragmatic, reproducible pattern for efficient code refactors and automation: leverage large-model context where it matters, and use lean CLI executors to act — a flexible template that’s especially useful for free or low‑budget AI-assisted development.