Why
The origin
I was pair-programming with Claude Code, building a small feature. At one point, I pushed back on its approach: "Use JWT for authentication, we need a stateless architecture."
Claude adjusted immediately. No friction. Great.
But later, I couldn't remember why I made that call. Neither could my teammate.
That's when it hit me: we have no way to persist human judgment in AI-assisted coding. The pushbacks, corrections, and redirections that make AI useful disappear when the session ends.
SteeringLog is my attempt to fix that.
What we believe
Steering is a growing skill
Steering AI is not a fixed trait. Like any skill, it can be developed through feedback and reflection. A record of your past decisions turns experience into material for improvement. Without it, growth is guesswork.
Team growth needs a shareable record
Good judgment might start as personal intuition, but with a record it becomes a shareable pattern.
Without a shareable record, each person learns alone, and the team's collective judgment stays fragmented.
Coding with AI changes how we measure performance
Code output is massive and hard to trace. Judgment reveals who has a genuine understanding of the conversation and who steers effectively. Making that visible turns an invisible skill into a measurable signal.
Token costs will become a heavy expense
Efficiency requires insight, but without records we optimise by blind guessing. Teams cannot see where repeated corrections or failed prompts burn tokens. Surfacing those patterns turns hidden waste into measurable savings.
AI's future needs fresh human judgment
Over-reliance on AI without reflection risks skill erosion, and training future models on AI-generated data leads them to collapse into stale loops. Real human judgment is the antidote. Keeping those signals in the training pipeline helps future AI stay grounded in authentic human artifacts.