The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting and processing arbitrary project documentation to create behavioral constraints for the AI agent.
The extracted content is stored in a canonical CONSTRAINTS.md file and the skill instructs the agent to treat these findings as non-negotiable boundaries, increasing the impact of any maliciously injected instructions found in the source documents.
Ingestion points: Automated discovery of repository files including docs/, README.md, decision records (adr/, rfcs/), and user-provided paths.
Boundary markers: Absent. The instructions do not require the agent to wrap or sanitize content from source documents when passing it to subagents or writing it to the findings file.
Capability inventory: File writing via scripts/merge_constraints.py and modification of existing behavioral files (AGENTS.md or CLAUDE.md) to establish a permanent context link to the extracted constraints.
Sanitization: The merge script performs basic normalization of metadata (titles and categories), but the narrative statements and impact descriptions are preserved without additional filtering or sanitization.
Metadata Discrepancy: The SKILL.md YAML metadata identifies the author as 'accelint', whereas the expected vendor context is 'gohypergiant'. While possibly a template remnant, mismatched metadata is a technique used to misrepresent the origin of a skill.
[COMMAND_EXECUTION]: The skill uses local shell commands and a Python script to manage its operations.
It executes bash loops for file discovery and uses python3 scripts/merge_constraints.py to merge data. These operations are limited to the local project environment and are consistent with the skill's primary function.

constraints-extractor