Deep Work Plan for developers and AI agents

Plan and execute with the skill

The API above lets an agent read this site. The DWP skill is what lets an agent run the methodology — install it once in a repository and it ships a router plus nine sub-skills, invoked as slash commands (or by name, for agents that intercept slash — most use # instead, e.g. #dwp-execute).

Two independent choices: format, and how much review you want

Every plan picks one value from each axis. They are independent — a Lite plan can run trust, a Full plan can run guided, and either can switch mode later without switching format.

The nine sub-skills

Each sub-skill is invoked as a slash command inside the repository that installed the skill — not against this website. The full reference for each lives in the kit catalog.

Two ways to run it

The same skill, the same nine commands — the format and review mode change with the size and stakes of the work, not the tool.

# A small, bounded fix: skip the review round, run it directly.
/dwp-create fix the flaky checkout test trust
/dwp-execute

# Long-horizon work with real stakes: review before anything runs.
/dwp-create migrate the billing service to the new payments API
# ...review the proposed plan, then:
/dwp-execute
# ...interrupted? pick up again, even in a fresh session:
/dwp-resume

Every plan’s output — manifest, progress log, task records, gate evidence — lives under a gitignored .dwp/ directory in the repository itself. Nothing is sent to or stored by deepworkplan.com; the skill makes no network calls at all.

Endpoints

The core endpoints of the agent API. The complete, typed specification — every operation, parameter, and response schema — lives at the OpenAPI document.

Unknown /api/* paths return a structured JSON error with a resolution hint, never an HTML error page.

Versioning & deprecation

The versioned JSON family lives under /api/v1/ — index, sections, pages, and health — and the unversioned canonical paths (/llms.txt, /{page}.md, /api/mcp) belong to the same v1 contract. Breaking changes ship only in a new /api/v{N+1}/ family, never inside v1. When an endpoint is deprecated, its responses carry Deprecation: true and a Sunset date at least 180 days before removal, and a Link header points at the successor.

Rate limits

Responses on /api/* carry RFC 9331 rate-limit headers — RateLimit-Limit, RateLimit-Remaining, RateLimit-Reset, and RateLimit-Policy — so agents can self-throttle in real time; a 429 response adds Retry-After. Enforcement is best-effort at the edge (120 requests per 60 seconds per visitor), and access stays anonymous: no keys, no registration, no tiers.

MCP server

A stateless Model Context Protocol server over Streamable HTTP. Three read-only tools: get_init_prompt, list_site_sections, and read_page. Protocol versions 2025-03-26 and 2025-06-18 are supported; no session is required.

# 1. Initialize (no session needed — the server is stateless)
curl -s https://deepworkplan.com/api/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-06-18","capabilities":{},
                 "clientInfo":{"name":"curl","version":"1.0"}}}'

# 2. Discover the tools
curl -s https://deepworkplan.com/api/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

# 3. Call one — read any page as source Markdown
curl -s https://deepworkplan.com/api/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call",
       "params":{"name":"read_page","arguments":{"path":"/init"}}}'

The MCP manifest lives at /.well-known/mcp.json and the server card at /.well-known/mcp/server-card.json. Claude, ChatGPT, and any MCP client can call these tools natively.

Markdown for agents

Every rendered page is published as native source Markdown — not an HTML conversion. Request Markdown explicitly with a URL suffix or through HTTP content negotiation on any page.

# Ask for Markdown on any URL (content negotiation)
curl -s https://deepworkplan.com/methodology \
  -H 'Accept: text/markdown'

# Or fetch the .md mirror directly — every page has one, in every language
curl -s https://deepworkplan.com/es/developers.md

Content negotiation returns the same source Markdown the site renders from, in the language of the URL you request.

Install the kit

The official install path for the Deep Work Plan skill — the same command the /init endpoint gives agents. It works with any skills-compatible coding agent (Claude Code, Cursor, Codex, Gemini, and others).

# 1. Install the DWP skill — same command the /init endpoint gives agents
npx skills add DailybotHQ/deepworkplan-skill@latest

# 2. Official CLI — zero-dependency client over this API (Node >= 18),
#    prepared in the site repo's cli/ directory pending npm publication
deepworkplan init
deepworkplan read /es/developers

The skill vendors into .agents/skills/deepworkplan/ inside your repository, so every agent that touches the repo shares the same methodology. The official deepworkplan CLI — a zero-dependency client over this same API (init, sections, read, open, mcp) — is prepared for npm and lives in the cli/ directory of the site repository until publication.

Machine-readable resources

• ▸ OpenAPI spec (/openapi.json)
• ▸ Agent access & auth declaration (/auth.md)
• ▸ API catalog, RFC 9727 (/.well-known/api-catalog)
• ▸ MCP manifest (/.well-known/mcp.json)
• ▸ Security contact (/.well-known/security.txt)
• ▸ Site repo descriptor (/.well-known/dwp.json)

Point an agent at it

The fastest path is still one line: hand any coding agent the /init prompt and it installs the skill, onboards your repository, and starts finishing deep work.