What is Deep Work Plan?

Deep Work Plan turns any repository into a structured environment where AI coding agents execute with precision and finish long-horizon work. It's not a chat window that forgets — it's a spec written directly into the repo: atomic tasks, acceptance criteria, validation gates, and resumable state. The plan becomes the durable source of truth, so any agent can pick up where the last one left off, even after context resets. Open Source under MIT, agent-agnostic, and built to eliminate drift in multi-hour coding tasks.

Who it's for

• Solo developers who want to walk away from a long-running agent task and return to verifiable, completed work without babysitting.
• Engineering teams running migrations, refactors, or new subsystems across dozens of files — where context loss and decision drift kill productivity.
• AI tooling enthusiasts experimenting with Claude Code, Cursor, Codex, or Gemini who need a portable, reusable harness that works across any agent without lock-in.

Key features

Spec-driven development with atomic tasks

The plan breaks work into atomic tasks, each with explicit acceptance criteria and validation gates. Agents execute against these specs, not against a fading chat history. Drift drops because the repository itself holds the standard.

Reasoning-based onboarding

When you point Deep Work Plan at a repository, it inspects the actual languages, frameworks, package managers, and CI commands — then generates artifacts adapted to that specific stack. A generic stub is treated as a failure. The result is AGENTS.md, a categorized docs/ tree, per-module READMEs, and a reasoned .agents/ kit.

Cross-agent portability

The harness lives in the repository, not in any single tool. A .agents/ directory with skills, agents, and commands works across Claude Code, Cursor, Codex, Gemini, and Copilot. The .claude to .agents symlink ensures every tool reads one source of truth — no duplication, no drift.

Resumable state across sessions

Long runs survive context resets. Any agent can resume where the last one left off, because the durable plan and its progress are stored in the gitignored .dwp/ folder. You hand an agent one line, walk away, and come back to work you can verify.

What stands out

"Models matter. Context matters more."

Deep Work Plan's core insight is that the best model is useless without a harness that keeps it on track. By embedding the plan, guardrails, and resumable state directly into the repository, it solves the fundamental problem of AI coding agents: they drift on long-horizon work. The result is a methodology that works with any agent, any repo, and no lock-in — battle-tested at Dailybot and released as open source.

Worth checking out if…

You're tired of watching AI agents forget what they were doing halfway through a migration, or you want to hand a complex refactor to an agent and trust it will finish without constant supervision. Deep Work Plan is also worth exploring if you manage multiple agents across different tools and need a single, portable harness that keeps everyone aligned to the same spec.