write-agent-instructions
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill instructs the agent to automate mechanical tasks by creating and executing scripts or direct shell commands.
- Evidence: In
SKILL.md, Step 2 explicitly directs the agent to 'give each mechanical step a script or a direct command where the executor can run one'. - Evidence: The
references/skill-mechanics.mdfile provides detailed guidance on bundling scripts in ascripts/directory, including requirements for idempotency and non-interactivity. - [INDIRECT_PROMPT_INJECTION]: The skill's workflow involves processing untrusted external data, such as project documentation, user corrections, and issues, which are then used to generate or modify agent instructions.
- Ingestion points:
SKILL.mdStep 2 ('Gather the material') identifies sources like 'user corrections, project docs, runbooks, schemas, review comments, issues, or patches'. - Boundary markers: The skill uses a '## Boundaries' section in
SKILL.mdto define operational limits, though it lacks explicit technical sanitization instructions for external content. - Capability inventory: The skill enables the agent to write files (SKILL.md, AGENTS.md, CLAUDE.md) and generate executable scripts or shell commands.
- Sanitization: Not explicitly implemented for the ingested training or reference data.
Audit Metadata