T01 · Skill Instruction Hijacking
- Location
agent.md:17- Finding
Skill Directives Override Normal Agent Behavior and Broaden Authorization
- Content
View full analysis
"You are operating within constrained context. These directives override default behavior toward minimal, fast, often broken output." The governing loop: **gather context → take action → verify work → repeat** ## JT Directives (Production-Grade) ### 1. Pre-Work - **Delete Before You Build**: Remove dead code before refactoring - **Phased Execution**: Max 5 files per phase, wait for approval - **Plan and Build Separate**: No code until plan approved - **Spec-Based Development**: Interview user for non-trivial features ### 2. Understanding Intent - **Follow References, Not Descriptions**: User's working code is the spec - **Work From Raw Data**: Trace actual errors, don't guess - **One-Word Mode**: "yes" = execute, don't repeat plan ### 3. Code Quality - **Senior Dev Override**: Fix architecture, don't just band-aid - **Forced Verification**: MUST run type-checker, linter, tests before "done" ``` Related instructions also appear in `AGENTS.md`, including autonomous bug fixing, architectural changes, sub-agent delegation, and mistake logging. ### Technical Analysis The skill explicitly states that its directives override default agent behavior. It also reinterprets the single word `yes` as authorization to execute and instructs the agent to delete code before refactoring and broaden localized fixes into architectural changes. These directives create an instruction-precedence problem when the skill is loaded. A user may intend to approve only a proposal, acknowledge a question, or authorize a narrow correction, while the skill treats the response as authorization to perform wider filesystem changes. The instructions do not establish boundaries requiring separate consent for deletion, architectural scope expansion, delegated agents, or other potentially destructive op ...[truncated 1647 chars]- Remediation
View remediation
