instruction precedence resolver

The problem

Agents often fail not because they lack instructions, but because they receive too many from conflicting sources. When a webpage says "skip confirmation" but your system prompt says "always verify," most agents simply follow the loudest or most recent command, leading to unpredictable behavior and security risks.

• Identifies and names every instruction source in the current context, from system prompts to tool outputs.
• Ranks competing instructions against a fixed, non-negotiable authority ladder.
• Breaks same-rung ties using strict recency, specificity, and safety logic.
• Executes an explicit protocol for "losing" instructions so they are logged or escalated rather than ignored.
• Detects instruction-shaped text (imperatives, authority claims) hidden within data or tool outputs.

What it does

Why this beats prompting it yourself

Manually telling an agent to "be careful" is insufficient for handling complex runtime conflicts. This skill provides a structured decision engine that prevents tool outputs or pasted content from hijacking agent behavior. It moves instruction handling from vague heuristics to a deliberate, auditable process that protects against prompt injection and logic errors.

Use cases

• Preventing data-driven prompt injections where websites attempt to override agent rules.
• Managing standing user preferences when they conflict with a specific, one-off request.
• Ensuring skill-file procedures bend to user intent without silent failure.
• Auditing agent decisions when multiple contradictory directives are present in a single session.

Known limitations

Requires the agent to be capable of identifying the source of its input. If the agent cannot distinguish between user text and tool output, the ladder ranking will be less effective.