Agentic skills framework that gates AI agent claims with evidence-based review
Provides a set of portable skills that fire at key moments in any agent harness to reconstruct intent, build evidence ledgers, and review work adversarially before claiming completion. It solves the common failure where agents drift from the original ask and report success invisibly. Works across Claude Code, Cursor, Codex, Copilot, OpenCode, Pi, and Gemini via harness-specific adapters that inject the same bootstrap. Useful for anyone building or relying on AI agents that produce code, documents, configs, or research.
View on GitHub →jinu1prakash/what-have-i-done