Context limits make long, multi-step agent work fragile: important decisions get lost, errors repeat, and tool-heavy sessions drift from original goals. This project treats the filesystem as working memory—writing plans, findings, and progress to disk so an agent can re-load authoritative state between tool calls and across sessions. That simple shift dramatically improves reliability for multi-step workflows.
What Sets It Apart
- Filesystem-as-memory pattern: enforces a three-file convention (plan, findings, progress) so the agent persists checkpoints and research rather than stuffing ephemeral context. This makes sessions recoverable and auditable.
- IDE & lifecycle hooks: integrates PreToolUse/PostToolUse/Stop hooks for Claude Code, CLI environments and other agents so the skill re-reads plans before actions and verifies completion before stopping—reducing goal drift and hidden failures.
- Evaluation-driven: includes formal evals (reported 96.7% pass rate on its benchmark) and many platform-specific fixes (Windows hooks, Codex/Copilot support) showing focus on robustness across real workflows.
Who It's For and Tradeoffs
Great fit if you run multi-step agent workflows, long-running research or multi-agent projects where context drift, session loss, or repeat failures are common. It’s especially useful when you want reproducible checkpoints and human-readable audit trails. Look elsewhere if your tasks are single-step lookups, tiny edits, or you need a fully managed cloud orchestration platform—this is a file-first, developer-focused skill that assumes you can run hooks in your IDE or agent runtime and accept files as the primary state carrier.
Where It Fits
Positioned between lightweight agent shortcuts and full orchestration stacks: it’s an operational pattern and skill that makes agents more predictable without adding heavy infra. Works well as the planning layer in agent stacks (RAG/agents/CI) or as a drop-in discipline for teams adopting Manus-style context engineering.
