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.
La règle la plus importante est la suivante : planifiez toujours en vous basant sur la vérité. Ne planifiez jamais sur la base d'une supposition ou d'un résumé. Un plan parfait construit sur des données obsolètes échouera aussi rapidement que l'absence totale de plan.
Communauté d'apprentissage optionnelle : https://t.me/GyaanSetuAi