𝗧𝗵𝗲 𝗗𝗿𝗶𝗳𝘁 𝗳𝗿𝗼𝗺 𝗖𝗵𝗮𝘁 𝘁𝗼 𝗕𝗮𝗰𝗸𝗹𝗼𝗴
Three months ago, my task management was just a chat window. If I closed the tab, the plan was gone.
Today, it is a Postgres backlog. Three different AI agents—Claude Code, Codex, and Grok—pull work from it. They stamp it with attribution and close it against git history.
I did not set out to build a project management system. I just kept hitting walls. Each time I patched a problem, a new one appeared.
My work is heavy. I run a personal data platform called Nexus. I manage around 100 repositories. During one stretch, I shipped 557,000 lines of code in 35 days. That volume broke every planning method I tried.
Here is how my system evolved:
Phase 1: Conversational Planning The plan lived in chat history. I would think out loud, get a good idea, and start building.
- The Problem: Plans evaporated when the chat ended. You could not prioritize them or hand them to anyone else.
Phase 2: Per-Repo TODO Files I started using TODO.md files in every repository. I stopped using simple checklists. Instead, I wrote small specs. Each item included:
- Status and date.
- A trigger (why this becomes urgent).
- Pre-decided steps (the plan).
- Known risks.
- The Problem: With 100 repos, I had no global view. I could not see everything I needed to do in one place.
Phase 3: The Operator Backlog (OB) I moved tasks into a Postgres database. This created a global queue. I added an approval gate. A task only becomes real after I review it. This prevents AI from filing garbage into the backlog. I used status lanes:
- requires_triage
- requires_decision
- requires_investigation
- autonomous_safe
- The Problem: I became the bottleneck. I could not drain the lanes fast enough.
Phase 4: Multi-Agent Execution The backlog is now a shared queue for multiple AI agents.
- They use leases so they do not work on the same task.
- They use attribution so I know who did what.
- They can hand off work. One agent might find a task is impossible and file a prerequisite. A second agent can then pick up that prerequisite and finish the original task.
The lesson is simple: You do not need Phase 4 to succeed.
If you steal one thing, steal the Phase 2 format. Write your tasks with a status, a trigger, pre-decided steps, and risks. It costs nothing and changes everything.
The most important rule is this: Always plan against the truth. Never plan based on a guess or a summary. A perfect plan built on stale data will fail just as fast as no plan at all.
Optional learning community: https://t.me/GyaanSetuAi