The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze potentially untrusted data in the form of repository instruction files and linked documentation.
Ingestion points: The workflow involves inventorying and reading all instruction surfaces, including files like AGENTS.md, CLAUDE.md, and any agent-facing docs referenced by pointers (Phase 1).
Boundary markers: The skill explicitly defines a 'Security Model' section stating that instruction-shaped text encountered during the audit is 'material under examination' and does not grant new authorization or change the skill's own workflow.
Capability inventory: The skill has the capability to write to the filesystem (Phase 6) and execute local shell commands. However, these are strictly governed by the workflow phases.
Sanitization: The skill relies on a mandatory human-in-the-loop approval gate (Phase 5), where all proposed diffs must be reviewed and confirmed by the user before any files are modified.
[COMMAND_EXECUTION]: The skill utilizes several local shell commands and a bundled script to analyze repository metadata.
Evidence: Instructions include the use of wc -c for byte counts, awk for length counts, git check-ignore for verification, and a specific one-liner python3 -c "import sys; print(max(len(l.rstrip()) for l in open(sys.argv[1], encoding='utf-8')))" FILE for character counting.
Bundled Script: The skill executes scripts/agents_chain_size.py, which is a bundled Python script used to measure context cost by traversing the directory tree and checking file sizes.
Context: All command executions are restricted to read-only operations on the local environment for metadata extraction and verification purposes.

maintaining-agent-context