Deep Work Plan

Deep Work Plan treats a written spec as the source of truth, not the chat log. You define the goal, scope, and acceptance criteria first, and the agent executes against them. The plan itself is the spec, so drift gets measured against stated criteria instead of your memory of what you asked for. That's what separates it from tool-bound systems that lock the spec inside one vendor.

Every task moves through a cycle of plan, atomic tasks, validation gates, resumable state, and completion. A task only closes when its checks pass, which means the agent can't simply declare the work finished and move on to the next item. Otherwise it stays open. That single rule removes most of the trust you'd otherwise need to extend to a machine doing engineering work while you're not watching. The loop is the same whether the job is a one-file fix or a multi-day build, which makes agent behavior predictable once you learn it.

Context lives in AGENTS.md and docs, tools in the .agents kit, guardrails in the validation gates, and state in a gitignored .dwp folder. Because every part is a file, the whole arrangement is portable and auditable. You can version it, review it, and hand it to a different agent without a rewrite.

Context is the scarcest resource an agent has, and DWP is engineered to spend it carefully because a careful approach that wastes that context will fail exactly when the work gets long and the stakes rise. Instructions load progressively as their trigger fires, and validation only touches what a task changed. The site publishes the instruction bytes each flow needs and reports that the latest release cut the execute flow by 45.7% and the resume flow by 67.5%.

A Lite plan covers a bounded fix, while a Full plan handles work that spans hours. The format follows the scope of the work rather than forcing every task into the same heavy template. Small jobs stay light. Big jobs get room for real structure.

The skill is a router over create, execute, refine, resume, status, verify, onboard, author, and upgrade. Each one has a thin slash command like dwp-create or dwp-verify. Refine lets you add, remove, or reorder tasks while preserving completed work, which matters when scope shifts mid-job.

Deep Work Plan ships adapters for Claude Code, Cursor, OpenAI Codex, GitHub Copilot, Google Gemini, OpenCode, Windsurf, Cline, and more. The adapters are thin by design, so updating your agent doesn't force a rewrite of your AI agent workflow. Swap tools. Keep the method. OpenClaw and Hermes get an unattended profile driven by heartbeat or cron scheduling.