How Structured Tracker Tasks Revealed Hidden Costs in AI-Assisted Development

Software engineer Anton, working on migrating a PHP monolith to Go services, observed that individual AI executor iterations were consuming around 350,000 tokens — far exceeding initial estimates. Investigation revealed the excess was driven by redundant context-gathering, where each executor independently rediscovered the same file paths, type names, and conventions without a shared record. To address this, Anton updated his specification format to include a closed list of files per iteration and a dedicated 'facts of the set' section, limiting file reads to at most three named files and prohibiting open tree searches. The core argument is that a tracker task must carry both a pre-work estimate and a post-work recorded fact, otherwise cost overruns remain invisible and project discussions devolve into subjective impressions. Anton notes the problem only became diagnosable because the task structure made the discrepancy measurable — turning what would have been a silent inefficiency into a documented spec defect with an attached cost.