Deep Work Plan – Turn a repo into a spec-driven harness for AI agents

Dailybot released Deep Work Plan, an open-source methodology that turns any repository into a spec-driven harness for AI coding agents, enabling long-horizon tasks with reduced drift. The tool generates repository-specific artifacts like AGENTS.md and docs/ by reasoning about the actual stack, and supports multiple agents including Claude Code, Cursor, and Codex.

Claude Code FullReference implementation, with native WebFetch and slash commands. Open methodology · MIT · Agent-agnostic Deep Work Plan turns any repository into a structured environment — context, guardrails, and a durable plan — where any coding agent executes with precision and finishes long-horizon work. Copy the init.md prompt and paste it into your coding agent — Claude Code, Cursor, Codex, or any other — to make any repository AI-first. Deep Work Plan is spec-driven development where the repository itself becomes the harness. The problem and the answer AI coding agents are remarkably effective in short bursts. On long-horizon work — a migration, a new subsystem, a refactor across dozens of files — they drift: context fills up, decisions are forgotten, and multi-hour tasks are abandoned halfway through. Deep Work Plan answers with spec-driven development: the plan is the durable source of truth, and agents execute against explicit acceptance criteria and validation gates. Drift drops, the work stays verifiable, and any agent can resume it across sessions. It is also harness engineering made portable. An agent harness is the scaffolding around a model — context, tools, control loop, guardrails, resumable state — that makes it reliable. Deep Work Plan installs that harness into the repository itself AGENTS.md, docs, the .agents/ skills home, the DWP skill , so any agent can pilot any repo. Born at Dailybot, battle-tested for months, and released as the DailybotHQ/deepworkplan-skill. Reasoning-based onboarding The onboarding flow inspects your repository's actual languages, frameworks, package manager, and validation commands, then generates artifacts adapted to that repository. A generic stub is treated as a failure. Reads manifests, folder layout, and CI to infer the real test, lint, and build commands, then classifies the repository as an individual repo or an orchestrator hub. A reasoned AGENTS.md, a categorized docs/ hierarchy, and a README plus docs/ inside each major module — filled with your repository's real commands, not placeholders. A cross-agent .agents/ directory skills, agents, commands and the .claude to .agents symlink, mirroring CLAUDE.md to AGENTS.md, so every tool reads one source of truth. Wires the Deep Work Plan skill and creates the gitignored .dwp/ folder for plans and drafts, then optionally layers opt-in addons such as devcontainer support. What happens when you run it You do not pick an install method or copy a template. You hand your agent one line; it installs the skill — the reusable engine — and adapts your repository to it. It reads the onboarding prompt at deepworkplan.com/init.md and the methodology, specification, and kit it links to — the standard it is about to adopt. The skill is the engine — the same in every repository. One command pulls in the router and its sub-skills create, execute, refine, resume, status, verify, onboard, author for Claude Code, Cursor, Codex, Gemini, and Copilot. Reasoning about your real stack — never copy-pasting — it writes AGENTS.md, a categorized docs/ tree, per-module READMEs, a reasoned .agents/ kit, and a gitignored .dwp/. Your repository becomes the harness. Generate long-horizon Deep Work Plans for any task and run them step by step, with explicit acceptance criteria, validation gates, and resumable state — autonomously, for hours. The skill is installed identically everywhere; what is adapted is your repository — the AGENTS.md, docs, and reasoned .agents/ kit generated for your stack. That split is what makes the methodology a reusable standard rather than a one-off scaffold. What you get One run, committed atomically. Every output is Markdown and every change is auditable. Reasoned from your repository's actual stack, commands, and structure — not a template with placeholders. CLAUDE.md is symlinked to AGENTS.md. Architecture, setup, standards, and troubleshooting — plus a README and docs/ inside each major module, generated from your codebase. A cross-agent .agents/ directory skills, agents, commands with the .claude to .agents symlink so every tool reads one source of truth. create, execute, refine, resume, status, verify, onboard, and author — available to your agent as a single skill pack, with no per-repository copy. /dwp-verify produces an objective pass/fail report against the specification, so "AI-first" is verified, not asserted — and re-verifiable after every plan. Onboarding classifies your repository as an individual repo the common case or an orchestrator hub that coordinates child plans across repositories. The author sub-skill skill-create, agent-create lets the repository evolve its own skills, agents, and commands; opt-in maintenance add-ons such as dependency-upgrade help it keep itself up to date. No daemon and no external state. Plans and drafts land in a gitignored .dwp/ folder, and any task resumes from git alone — even after context overflows. Agents One methodology, many adapters. Markdown couples the framework to nothing — every agent that reads Markdown can run a Deep Work Plan. Reference implementation, with native WebFetch and slash commands. Full adapter. Use the offline bundle if WebFetch is gated. Offline bundle recommended; rules installed under .codex/. Full adapter — the dwp- commands run via AGENTS.md and procedures. Requires Gemini 2.5 Pro or newer, with native WebFetch. Open source. Reads AGENTS.md natively and runs dwp- via commands. Rules plus command procedures drive the full Deep Work Plan loop. Open source. Markdown rules and commands run every dwp- step. Full adapter with a native command surface. Stacks These are reasoning aids, not templates. Onboarding reads your repository's real manifests and adapts per stack — it never blind-copies a preset. Monorepos get per-module docs. Two archetypes Onboarding forks on the archetype. Most repositories are individual repos. A hub coordinates child Deep Work Plans across many repositories. The methodology handles both as first-class. A single codebase with one primary stack, its own validation commands, and per-module docs. The default — onboarding assumes it unless the repository is clearly a hub. For example, a Django API, a Vue app, or a TypeScript Lambda service. A coordination repository that orchestrates work across sub-repositories via an orchestrator manifest, spawning child plans that each commit in their own repository, plus boundary rules and a navigation index. For example, a hub coordinating five product repositories. Methodology versus tool Deep Work Plan is not another scaffolder. It is the methodology layer underneath any spec-driven or scaffolding tool, focused on multi-hour autonomous runs. | Methodology versus tool | Deep Work Plan | Scaffolding / spec tools | |---|---|---| | Primary focus | Multi-hour autonomous execution | Spec or scaffold generation | | Unit of work | A Deep Work Plan resumable session | A spec document or a scaffold | | State model | Git-native .dwp/ folder, resumable | Often external or in-IDE | | Agent coupling | Agent-agnostic Markdown and Bash | Often tool- or IDE-specific | | Context recovery | Resumes after context overflow | Typically restarts the task | | License | MIT, open methodology and kit | Varies | Make your repository AI-first Hand your agent one line — point it at /init.md — and it makes your repository AI-first: it installs the skill, reasons about your stack, and commits a complete AGENTS.md hierarchy. From there you create and execute Deep Work Plans that run autonomously for hours.