The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from the repository being analyzed, including README files, source code comments, and git commit messages. This creates a surface for indirect prompt injection if the analyzed repository contains malicious instructions. The skill mitigates this risk by treating extracted signals only as candidates for a human-guided interview process and requiring multiple sources of evidence before accepting a capability specification.
Ingestion points: README.md, spec/charter.md, spec/system-map.md, CLAUDE.md, AGENTS.md, source code directories, script files, documentation, and git commit history.
Boundary markers: The skill instructions explicitly define extracted data as "interview seeds" and "raw signals" that must be validated through a per-capability interview flow and a 3-axis predicate test.
Capability inventory: The skill uses fs.readFileSync for file access and child_process.execFileSync for running git commands.
Sanitization: Input strings for identifiers are processed through a slugifyCandidate function that cleans and normalizes the text.
[COMMAND_EXECUTION]: The skill executes the system git command to extract recent commit history, which is used to identify active work areas in the repository.
Evidence: scripts/extract-signals-core.js uses execFileSync('git', ['-C', repoRoot, 'log', ...]) to gather commit metadata.

spec-grill