convention-injector

To inject task-specific conventions into subagents and prevent prompt contamination, use a skill that reads user and project CLAUDE.md files, filters relevant rules, and appends a compact conventions block with a self-confirmation suffix to the subagent prompt.

Prompt contamination occurs when subagents execute tasks with irrelevant or conflicting instructions. CLAUDE.md rules prevent it by enabling task-specific filtering, ensuring subagents only receive applicable conventions within a strict token limit.

To apply project conventions when spawning Explore and Plan agents, run the convention injection process before spawning to load CLAUDE.md, filter relevant rules, and inject a compact conventions block into the agent prompt.

Yes, convention injection works with user-level and project-level rules simultaneously by collecting conventions from both user and project CLAUDE.md files, as well as rules/*.md, to filter and apply relevant rules to the subagent.

The token limit for injected conventions in subagent prompts is 300 tokens. The injection mechanism filters relevant rules to create a compact conventions block that stays within this strict limit to prevent prompt bloat.

To add a self-confirmation prompt for audit-type subagent tasks, use the anti-contamination option provided by the convention injector, which appends a self-confirmation suffix to the subagent prompt to verify rule adherence.