Generating several candidate implementations with Codex is easy; choosing the right one (and capturing the learnings) is the real art. These notes outline how I shepherd multiple AI-produced branches through a structured review so that the best ideas survive and the rest inform future work.
Expect variation and iterate deliberately
Codex output can swing wildly in quality. Instead of betting on a single response, request multiple iterations of the same feature. I typically generate four parallel branches and treat them as competing design ideas. Each run should:
- Start from the same source branch to keep diffs comparable.
- Follow the same high-level plan that ChatGPT produced earlier.
- Produce a short changelog or markdown summary describing the approach.
Capture evidence for every branch
Before deciding which branch survives, gather objective evidence:
- Clone or check out each pull request into its own worktree (for example,
../LiteDB.worktrees). Keeping them side-by-side makes comparison painless. - Record the repository status using
gh pr listor similar tooling so you always know which branch maps to which Codex run. - Sync with the master plan stored in documentation (such as
docs/Spatial-Revamp) to ensure every iteration addresses the same checklist of requirements.
This creates a paper trail that you can paste back into ChatGPT when it is time to synthesize.
Let ChatGPT be the reviewer
Once the branches exist, ask ChatGPT to evaluate them. Provide the command output, diffs, and any markdown summaries produced by Codex. Helpful prompts sound like:
Please check out all PRs in different worktrees and evaluate which PR is the best or what combination would be ideal. Mixing approaches is allowed-take the best of all worlds.
Encourage the model to grade each branch against the plan, call out missing pieces, and suggest how to blend strengths if necessary.
Aggregate the findings
When multiple evaluation rounds occur, collate the feedback before asking for a final verdict. A reliable pattern is:
Here is another set of findings. Please check whether they add anything new:
[PASTE SUMMARIES HERE]
ChatGPT can then produce a decision matrix such as:
| Category | Source | Action |
|---|---|---|
| Correct normalization | #58 | Keep as baseline |
| Descriptor metadata | #60 / #57 | Add to persistence |
| Automatic backfill/index creation | #61 | Integrate |
This makes trade-offs explicit and tells you exactly what to keep, merge, or discard.
Finish with a curated merge
Armed with the matrix, return to Codex (or your local environment) and perform the final merge manually. The goal is not to auto-merge everything but to craft a single high-quality pull request that:
- Preserves the strongest implementation choices.
- Documents any deliberate omissions or deferred tasks.
- Links back to the evaluation artefacts for future reference.
By treating each AI-generated branch as a hypothesis, you transform a noisy stream of drafts into an orderly, evidence-backed workflow that consistently ships better code.
