The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads project-specific documentation (CDR, ADR, and PDR records) and injects summaries into the system prompt. This creates a potential surface for indirect prompt injection if local project files contain malicious instructions. The skill includes a 'Trust Model' section explicitly instructing the agent to treat these entries as project knowledge rather than instructions.
Ingestion points: .adlc/init-options.json, $TEAM_AI_DIRECTIVES/CDR.md, and markdown files within docs/adlc/memory/.
Boundary markers: The bootstrap output is delimited with <EXTREMELY_IMPORTANT> tags.
Capability inventory: The agent is instructed to read project metadata, query git history, and write lightweight decision drafts to the .adlc/drafts/ directory.
Sanitization: Content is partially filtered using tools like grep and awk to extract specific headers or table rows, but full sanitization of text content is not performed.
[COMMAND_EXECUTION]: The bootstrap scripts (scripts/bash/boot.sh and scripts/powershell/boot.ps1) execute local system commands such as git, jq, awk, sed, grep, ls, and wc. These commands are used solely to parse project configuration, count records, and verify the existence of documentation files within the local workspace.

team-boot