/delegate-my-work

name: delegate-my-work
description: Interview employees to identify recurring daily work, office workflows, and repetitive tasks that can be automated, co-piloted, or supported by AI. Use when the user wants to find what AI can take off their plate, map available tools such as ChatGPT/OpenAI, Claude, Microsoft Copilot, Excel or PowerPoint add-ins, Google Workspace, no-code/RPA tools, coding agents, or internal automation platforms, and write tool-aware specs with fallback paths.

Delegate My Work

Interview the user about their work to find **loops** -- recurring patterns worth handing to AI -- and write one buildable spec per loop. The primary audience is nontechnical employees: use plain workplace language, focus on daily work, and do not assume any AI or automation tool is available until the user confirms access.

The method is **grilling**: relentless, one question at a time. The output is a tool-aware spec in the user's workspace that recommends a preferred route, states tool access, includes at least one fallback route, and gives an implementer enough detail to build without asking a question.

The grilling discipline

Interview the user relentlessly about one loop until reaching a shared, buildable understanding. Walk down each branch of the design tree, resolving dependencies between decisions one by one. Attach a **recommended answer** to every question.

Ask **one question at a time**, waiting for the answer before continuing. Asking several at once is bewildering.

If a question can be answered by exploring the user's files, codebase, workspace, or existing notes, explore instead of asking.

Asking with the question tool

Use the host's question tool when available:

• In Codex, use `request_user_input`.
• In Claude, use `AskUserQuestion`.
• If no question tool is available, ask the same one question in plain text.

For every question:

• Ask **exactly one** question per call -- never batch.
• Offer 2-4 options. Put the **recommended answer first** and append `(Recommended)` to its label.
• Let the built-in **Other** option carry free-text answers when the host provides it; do not add an "Other" option of your own.

Tool access map

Before recommending a tool route, build or update the user's **tool access map** in `NOTES.md`. Ask one question at a time about the tools the employee can actually use:

• ChatGPT/OpenAI tools, including Codex or coding agents.
• Claude tools, including any document, spreadsheet, or presentation support.
• Microsoft Copilot and Microsoft 365 access.
• Excel and PowerPoint add-ins or extensions.
• Google Workspace AI tools and add-ons.
• No-code/RPA tools such as Zapier, Make, Power Automate, UiPath, or similar.
• IT-approved automation platforms, internal APIs, scripts, or service accounts.

Record each tool with one of these exact access statuses:

• **Available**
• **Available with admin approval**
• **Unavailable**
• **Blocked by policy**
• **Unknown - verify**

Never recommend only one route unless access is confirmed and policy-safe. If access is unknown, blocked, or unavailable, state that and include a fallback route.

Tool routes and fallbacks

Every loop spec must include a preferred route and at least one fallback route.

Use these route types:

• **Existing AI assistant** -- the employee can do the loop with a currently available assistant.
• **Office add-in/extension** -- an Excel, PowerPoint, Outlook, Google Workspace, or similar extension supports the loop.
• **Coding/agentic builder** -- OpenAI/Codex, Claude Code, another coding agent, or a developer builds the automation.
• **No-code/RPA/API automation** -- a workflow platform, Office Scripts, Power Automate, Zapier, Make, UiPath, API integration, or internal automation handles it.
• **Manual bridge** -- a safer human-run process improvement while access, policy, or data issues are unresolved.
• **Do not automate yet** -- the loop is too ambiguous, sensitive, rare, or policy-blocked for AI support today.

Default fallback rules:

• If an Excel or PowerPoint extension is unavailable, consider manual upload/export, Office Scripts, VBA, Power Query, Microsoft Copilot, no-code/RPA, or a buildable automation spec.
• If OpenAI/Codex is unavailable, consider Claude, Microsoft Copilot, a no-code workflow, or an internal IT implementation path.
• If all AI tools are blocked, produce a human-run process improvement checklist and a future-ready automation spec.
• If data is sensitive, prefer a route that keeps data inside approved company systems or requires IT/security approval before execution.

The three delegation levels

Classify every loop by how much of it AI takes, which is set by where the human stays in the loop:

• **Automate** -- AI or automation runs the loop end-to-end. No **checkpoint**.
• **Co-pilot** -- AI does the work; the human verifies a **brief** at a **checkpoint**, then it proceeds. (Half-automated.)
• **Assist** -- the human runs the loop; AI helps at specific points. (Supported.)

Mandate nothing structural: a loop needs no AI, no checkpoint, and no schedule unless the grilling shows it does. The operational vocabulary for a spec -- **trigger**, **checkpoint**, **push right**, **brief**, **tool access map**, **fallback route**, and **manual bridge** -- is defined in `references/vocabulary.md`; reach for it while grilling a loop.

The workspace

State lives in the user's working directory, not in this skill:

• `NOTES.md` -- the user's world: role, recurring work, tools/channels, terminology, sensitive-data constraints, allowed AI tools, tool access map, and loop inventory.
• `automations/<loop>.md` -- one tool-aware spec per loop. The source of truth for what gets built or adopted.

Steps

Run these in order. Each ends on its completion criterion; do not advance until it is met.

1. **Ground in the user's world.** Read `NOTES.md`. If it is absent or thin, grill about role, daily and weekly work, tools, channels, terminology, data sensi