I grew my agent instruction file from 90KB to 147KB. Violations did not drop. The 8 rules I moved into code did.

Описание и применение

Увеличение размера инструкции агента с 90 КБ до 147 КБ не снизило количество нарушений правил. Перемещение восьми наиболее часто нарушаемых правил в кодовые гейты привело к нулевому количеству нарушений, в то время как оставшиеся 16 правил, оставшиеся в документе, продолжали нарушаться. Большая часть токенов использовалась на повторное чтение инструкции, что делает суммарные токены плохой метрикой работы.

Задача

Increasing the size of an agent’s instruction file did not reduce rule violations, and the agent continued to ignore documented rules.

Результат

Moving the most‑violated rules into code gates eliminated their violations, while expanding the instruction file did not reduce overall violations. Token usage was dominated by re‑reads of the instruction file, making cumulative tokens a poor metric of work.

Рабочий процесс

• 1. Measure token usage and violation counts over 48 days.
• 2. Split 24 rules into two groups: 8 most‑violated (Group A) and 16 others (Group B).
• 3. Move Group A rules into code gates (pre‑tool‑use hooks, checker scripts, guard before send).
• 4. Keep Group B rules as document clauses but inject only the top 3 most‑violated clauses each turn.
• 5. Log violations, track learning queue, and apply a tiered rule‑creation policy (log first violation, write clause on second, code gate on third).
• 6. Monitor recurrence rates and token re‑reads to evaluate effectiveness.

Доказательства

94.4% of my tokens were re-reading the same text

Group A: 8 rules, 0 recurrences; Group B: 16 rules, 19 recurrences

Rejection rate 49.8%