Backlog.md - Reaching 95 % Task success rate with AI Agents

Abstract

I started by throwing Claude at my codebase and spent more time fixing its “help” than shipping features. Few months later I perfected the process and built Backlog.md. Now tasks are markdown files in my repo and Agents are using them effectively and I can finally build all the side projects that I have been accumulated as dormant domains.
This talk will take the audience from what would a complete beginner do with AI Coding assistants to a master level and achieve close to 100% success rate with their tasks.

The deck will be refreshed with best practices right up until October 2025.

Outline (35 min)
Why the first attempts flopped – “naked” prompts and repo chaos (5 min)

1. 50 % → 75 % → 95 % – adding just enough context files (15 min)
2. Backlog.md CLI demo – turning a sentence into parallelizable tasks (5 min)
3. Claude vs. Codex vs Jules: strengths, weaknesses, peace treaty (5 min)
4. Phone‑only workflow & live merge (5 min)

Audience take‑aways

• A drop‑in folder structure that makes LLM code contributions safe.
• A repeatable loop to churn through features on autopilot.
• A checklist to decide which agent should plan, implement, or review.

(And yes, the checklist will grow as the tools do—see you in October.)