The Agent Skills Directory

[SAFE]: The skill follows secure development practices by utilizing explicit user prompts for important decisions and validating the structure of generated artifacts using a local script. It does not attempt to access sensitive system files or credentials. All bundled scripts are read-only templates and the process is designed to be agnostic to the host project.
[COMMAND_EXECUTION]: The skill invokes several local commands to manage the review workflow:
It executes git status to monitor for unexpected filesystem changes during the review process.
It runs a bundled shell script scripts/validate-findings.sh to ensure the integrity of the findings report. This script uses standard utilities like grep, sed, and awk for text validation.
It calls the compozy binary to interface with external LLMs. Input parameters like --ide and --reasoning are validated against predefined sets of allowed values in SKILL.md.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes external specification files for review.
Ingestion points: The contents of files targeted by spec-path and --context (in SKILL.md) are ingested into the context for the reviewer LLM.
Boundary markers: The skill uses structured templates with specific headers (e.g., CONTEXT FILES TO READ:, SCOPED-WRITE CONTRACT:) in references/peer-review-prompt.md to separate instructions from data, which helps mitigate instruction confusion.
Capability inventory: The skill can write files to the local directory, execute local shell scripts, and invoke the compozy tool to send data to an external LLM provider (in SKILL.md).
Sanitization: No content filtering or escaping is applied to the specification data before it is interpolated into the prompt.

spec-peer-review