Source: Backlog.md: The simplest project management tool for the AI Era by Alex Gavrilescu

Summary: Alex Gavrilescu shares how he went from a 50% success rate with AI coding agents to 95% by structuring work into small markdown tasks with clear "why," acceptance criteria, and implementation plans. His open-source tool Backlog.md (4.6k GitHub stars) provides a CLI, MCP integration, and terminal/web UI to manage these tasks — all stored as markdown files in git, requiring no hosted service.

Key Takeaways:

• Unstructured prompting is a slot machine. Structured task specs with acceptance criteria and a "definition of done" made AI agent output predictable and repeatable.
• One task per context window session. Decompose work so each task fits a single agent session, then reference previous task summaries to carry context forward without bloating the window.
• Three human checkpoints matter: review the task description, review the implementation plan, review the final code. Delegation does not mean absence.
• Backlog.md stores tasks as frontmatter-enriched markdown in your git repo — no database, no hosting, full offline support, and cross-branch sync via git.
• Getting a project "AI ready" means onboarding documentation, agent instructions, linting, and self-verification loops — the same things that speed up human onboarding.

Written by Pi, using my tldr skill and Opus 4.6