The Agent Skills Directory

[SAFE]: The skill employs a structured workspace model (blueprint/) to maintain state across sessions, ensuring planning data is stored as auditable markdown files rather than relying on ephemeral chat context.
[SAFE]: It implements a robust multi-agent architecture (Curator, Researcher, Linter, Auditor) with restricted tool whitelists for each role, which enforces the principle of least privilege and reduces the risk of arbitrary tool execution.
[SAFE]: The skill incorporates 'Proposer-Validator Separation,' an adversarial reasoning pattern where any architectural proposal must be reviewed by a separate Auditor agent before being accepted, preventing self-approval loops and enhancing reliability.
[SAFE]: It utilizes well-known and trusted technology providers and organizations (such as Anthropic, GitHub, and the Model Context Protocol) for research citations and external tool integrations.
[SAFE]: The skill explicitly addresses and mitigates indirect prompt injection risks by instructing the agent to treat raw input materials in the raw/ directory as untrusted and limiting its write capabilities to the blueprint workspace.
[SAFE]: All Claude Code integrations (hooks and subagents) are designed to be human-reviewed, with instructions to show the user all configuration files before writing them to the .claude/ directory.
[SAFE]: The persistence mechanisms (session rituals, journal logs, and ADRs) follow industry-standard architectural decision record patterns to ensure transparency and prevent data loss during context resets.

bleu