Sixteen terms that show up in every Claude Cowork conversation, defined the way Tinkso uses them in client work. Anchor IDs for deep-linking from outside.
A short, opinionated dictionary of the words that show up in Cowork conversations. We define them the way Tinkso uses them in client work — not the way every blog post does. Each entry includes the term, a two-sentence definition, and a link to where it lives in the bible. Anchor IDs are stable, so you can link directly from email, decks, or other docs (for example /bible/foundations/glossary#mcp-connector).
A Claude run that plans, takes actions, and reports back, rather than just answering. In Cowork, every multi-step prompt is an agent run. See: What is Cowork.
The user-confirmation step Cowork inserts before destructive or ambiguous actions. Tinkso recommends keeping approval gates on for the first 30 days of any deployment. See: First 30 minutes.
A markdown file in the workspace folder that Cowork reads on every session to learn house rules: tone, file naming, formats, things never to do. The shared written contract that survives memory resets. See: CLAUDE.md and memory.
A small software bridge that lets Cowork talk to a SaaS app (Slack, GitHub, Notion, Asana, Salesforce, etc.) using the Model Context Protocol. Some are first-party from Anthropic; others are built per engagement. See: MCP connectors.
The size of the prompt and history Cowork can hold in a single run. Sonnet 4.6 and Opus 4.6 support up to 1M tokens (in beta) — roughly 1,500 pages of material. See: Models and context.
The middle tab in the Claude desktop app. Gives Claude file access to one folder you grant it. Different from Chat (no files) and Code (developer-oriented).
Cowork's persistent storage of facts across sessions. Memory survives restarts; it does not survive account changes. Per-user, not per-workspace.
The open spec Anthropic uses for connectors. If a SaaS has an MCP server, Cowork can talk to it. See: the marketing-side primer at What is MCP.
Cowork's pre-action step where it lists what it will do, then waits for approval. Skipping plan-mode is the most common cause of Cowork "did the wrong thing" complaints.
A bundle of skills, prompts, and connector configs that Claude Code or Cowork can install in one step. Tinkso ships custom plugins per client engagement.
A natural-language instruction. In this bible, "prompt" includes any specifying detail (folder paths, formats, constraints) — not just the question. See: Prompt anatomy.
The rolling envelope of model usage on a Cowork seat. Resets every five hours by default. See: Usage limits.
Anthropic's term for "still cooking." Cowork left research preview at GA on April 9, 2026; some sub-features (memory, scheduled tasks) shipped after.
A reusable instruction set Cowork can invoke (for example, a "month-end close" skill). Skills are how you scale a successful prompt across a team. See: Skills and plugins.
A Cowork run launched by another run for a sub-task — for example, "summarise each file individually before combining." Useful for long jobs and for keeping the parent run focused.
The single folder you grant Cowork access to. Tinkso's pattern is one workspace per function, with inbox/, output/, and archive/ inside. See: The workspace pattern.
Book a 30-minute call. We'll ask where you are, what your team needs, and which systems Cowork should touch.