PAAD — Defense-in-Depth for AI-Assisted Development

PAAD is a Claude Code plugin marketplace and multi-platform skill suite that adds defense-in-depth safeguards to AI-assisted development. It catches failure modes AI coding assistants don't reliably catch on their own: weak specs, plan-spec drift, architectural decay, and skipped quality gates. It supports Claude Code natively and ships compatible skill files for Cursor, Kiro, and Antigravity.

Eight skills form the suite. Pushback critiques specs before implementation begins, checking for contradictions, feasibility, scope problems, omissions, ambiguity, and security concerns — one issue at a time to respect developer attention. Alignment checks that requirements, plans, and tasks are consistent, then rewrites all action items in TDD format with specific red/green/refactor steps. Agentic-architecture dispatches five specialist agents in parallel, each examining the codebase through a distinct architectural lens across 34 flaw types and 14 strength categories. Fix-architecture guides iterative, sequential fixing of those flaws with mandatory safety-net tests before any refactoring begins. Agentic-review is a pre-merge quality gate using five bug-hunting specialists covering logic, error handling, contract compliance, concurrency, and security. Agentic-a11y audits accessibility across eight platform types with platform-specific checks per specialist, targeting WCAG 2.2 AA. Vibe guards small fixes with mandatory TDD pre-flight checks. Makefile generates and updates project Makefiles with strict rules against modifying existing targets without approval.

The core architectural pattern across the multi-agent skills is specialists-then-verifier: dispatch parallel agents with distinct lenses, then have a single skeptical verifier read actual code at every referenced file:line, drop false positives, deduplicate findings, and assign severity. This treats the LLM's tendency to generate plausible-but-wrong claims as a solvable reliability problem rather than a given.

Every skill except help includes a Graphviz decision digraph encoding its safety gates visually, making it harder for the LLM to skip pre-flight checks. The build system enforces this and five other invariants via Makefile targets. Skills are invoked as /paad:<name> in Claude Code and via natural language on other platforms. A conversion script strips Claude Code-specific paths and references for the Kiro and Antigravity ports.

The philosophy is honest about risk, keeps humans in the loop for every consequential decision, and optimizes for better outcomes over minimum token consumption.