𝗧𝘂𝗿𝗻 𝗬𝗼𝘂𝗿 𝗣𝗹𝗮𝗻𝘀 𝗜𝗻𝘁𝗼 𝗗𝗮𝘁𝗮
We spent months in a review loop. LLM agents planned and reviewed code. Work kept bouncing back. Every bounce cost time.
We analyzed 2,400 review documents. We found 6 reasons for failure. Most were not code bugs. They were plan defects. Missing docs. Unstated rules. Gaps in proof.
Prose plans do not bind code. Reviewers fail to rerun a text plan. We changed the format. We used TOML. We turned plans into data.
We built a Python validator. It checks the plan before a human sees it.
- It finds cycles.
- It checks dependencies.
- It ensures every claim has proof.
- It rejects placeholders.
A wrong plan is now a failed assertion. The human reviewer does not waste time on structure. They focus on domain risks.
Our first DAG-reviewed plan passed in one round. Reviewers spent their budget on real risks.
Your review archive is a dataset. Use it to find patterns. Stop using prose for technical plans. Use data.
Source: https://dev.to/wernerk_au/dag-toml-how-we-turned-four-months-of-code-review-pain-into-a-machine-checkable-planning-format-236j Optional learning community: https://t.me/GyaanSetuAi