Overview
The Setup Cowork system prompt guides users through a six-step progressive onboarding flow to configure Cowork, an AI assistant that autonomously handles tasks like reading email, searching docs, and drafting reports. The prompt teaches reusable workflows (Skills), connectors, and plugins through methodical steps rather than overwhelming the user all at once.
Step-by-Step Flow
Each step in the flow includes explicit guardrails against presuming tool results before widgets run, always calling the tool first and reacting to its output. The flow separates tool configuration (plugins, connectors) from user personalization (writing voice in Step 5).
Key Guardrails
- Never presume tool results — Each step forbids writing text that assumes widget output before the tool runs; always call the tool first, then react to its output.
- Tool vs. user separation — Steps 1–4 configure tools; Step 5 teaches the AI the user’s writing style from previously sent writing, but only if a profile doesn’t already exist.
- Compounding invariant — Step 6’s wrap clause ensures that if no voice profile is established during setup, the user can always create one later by asking, preserving the “preserve and extend” invariant.
Uncertainties & Conflicts
- Missing tool fallbacks — The prompt references tools (ShowOnboardingRolePicker, ListPlugins, SearchPlugins) that may not be available in all sessions; the skill says to skip if unavailable, but fallback behavior when multiple tools are missing is unspecified.
- Writing voice ambiguity — Step 5 warns against re-running the flow on someone who’s already set up a profile, yet also says a recently saved profile may not show in the skills list yet, creating ambiguity about how to check for existing profiles.
- Step resumption after skill invocation — The ‘one step at a time’ ground rule conflicts with users trying a skill mid-flow; the prompt says to help then return, but doesn’t specify how to track which step to resume from.
- ‘Loosely relevant’ definition undefined — The org-plugin priority rule says plugins ‘even loosely relevant to the role outrank any generic one, full stop,’ but ‘loosely relevant’ is undefined.
Implications
This prompt exemplifies a progressive-onboarding pattern that could be generalized across other AI assistant setup flows. The explicit guardrails against presuming tool results and the separation of tool configuration from user personalization are transferable patterns. The compounding invariant in the wrap clause (Step 6 fallback) is a reusable design decision for ensuring users can always extend their assistant’s capabilities later.
Open Questions
- How should an AI assistant check for an existing writing voice profile when the skills list is plugin-filtered and may not show recently saved profiles?
- What constitutes ‘loosely relevant’ when prioritizing organization plugins over generic marketplace plugins?
- When multiple referenced tools are unavailable, what is the appropriate fallback behavior — skip the step entirely, use a text-based alternative, or proceed without the tool?
- Can the progressive-onboarding pattern be adapted for non-Cowork assistants, and what elements are essential vs. assistant-specific?
Relationships
- “implements” cowork-onboarding-pattern — the Setup Cowork system prompt implements this progressive onboarding pattern
- “is-sibling-of” anthropic-research-instructions — both are leaked system prompts with guardrails and structured flows
- “references” cowork-onboarding-pattern — evaluates the onboarding pattern implemented by this prompt
- “references” product-compass-pm-pattern — Product Compass PM newsletter pattern; overlaps with onboarding patterns
Sources
- raw/prompts/leaked/setup-cowork.md