Split thinking from editing: ChatGPT designs, coding agents execute — with 55% faster runs

fykup · reddit · 2026-08-27

The author advocates a "separate thinking from editing" workflow: architecture, design, planning, and independent review go to a high-reasoning ChatGPT session; filesystem-heavy execution (editing, tests, fixes) goes to Codex/Claude Code/OpenCode.

Since ChatGPT doesn't live in the repo, the author built the open-source, local-first AI Badger: instead of stuffing context, it packages repo topology, Git state, task context, and relevant source; ChatGPT can then request specific files/classes/symbols on demand. Once design settles, a compact implementation handoff goes to the coding agent: repo context → ChatGPT reasoning → compact plan → agent execution. An Agent Skill also carries over session state from coding CLIs into ChatGPT.

In one OpenSource dogfooding run, a compact Badger-assisted handoff vs. a raw feature request yielded 32% fewer agent tokens, 55% lower runtime, and 86% fewer reasoning tokens — one run, not a benchmark, but directionally consistent with daily use.

GitHub: github.com/PVRLabs/aibadger

Original post →

More from coding & agent

coding & agent channel →