Writer’s Brief

The Hook

What if every decision, bug fix, and midnight build session was training data for a PM agent that could eventually run your projects?

Main Points to Cover

  1. The idea: a meta-log that watches everything happening across your task system and extracts “story arcs” - not just events, but narratives with lessons
  2. The build itself was a perfect example of the process: brainstorm 3 options, pick the ambitious one, hit a stale-data bug, fix it live. The meta-log captured its own creation.
  3. The practical output: 693 events backfilled from git history, task threads, execution logs, and memory files. 7 story arcs auto-detected. A training export ready for fine-tuning.

The Angle

Personal discovery / building in public. You built a project management system, then built a system to learn from the project management system. It’s turtles all the way down, but the turtles are useful.

Target Reader

People building with AI agents who want to go beyond “AI does tasks” to “AI learns how I work.” Builders who are using Claude Code, Cursor, or similar tools and want to capture the meta-patterns.

Tone Notes

Casual, nerdy-excited. The absurdity of building a system that watches you build systems is part of the fun. Lean into the recursive quality without being precious about it.

Raw Material / Moments to Write From

These are notes for Alex to write from. Not prose.

  • Alex’s original ask: “a log that tracks everything we’re doing and updating with a goal of learning in real time what’s happening, how it all works. so i can train a PM agent with it all?”
  • The constraint that made it work: “I would love to do C but have it be a separate build that doesn’t affect the current mission control at all.” Separation was the unlock. No coupling, no risk, just observation.
  • The brainstorm presented 3 options ranging from a 20-line append to a full story arc system. Alex picked the most ambitious one immediately. Pattern: when given curated options with clear tradeoffs, the decision is instant.
  • First version showed stale data. Alex: “I’m not seeing any of the latest stuff? and it looks like 4 tasks are stuck?” This is the classic static-file trap. Fixed by building a live API server. The fix itself became an event in the meta-log.
  • 693 events extracted from data that already existed. Every git commit, every task thread message, every execution log. The history was already there, it just wasn’t connected into a story.
  • 7 story arcs auto-detected: Board UI evolution, C4C cohort ops, CLI for agent access, email pipeline, ITW podcast launches, queue runner removal, subtask addition. Each one is a “season” of the project.
  • The recursive quality: the meta-log’s own creation (brainstorm, build, debug, ship) is itself a story arc in the meta-log. It captured itself being born.
  • The PM training angle: the training-export.jsonl is structured for fine-tuning. Each entry has the event, the context, and the lesson. The idea is to feed this to an agent that learns not just what happened, but why decisions were made.
  • Total build time from brainstorm to “works!!”: about 2 hours across a few sessions. Most of that was the backfill logic parsing 5 different data formats.