> ## Documentation Index > Fetch the complete documentation index at: https://docs.agentdesk.team/llms.txt > Use this file to discover all available pages before exploring further. # Set up a new workspace > A checklist to make a new workspace ready: providers, members, projects, agents, rules, memory, skills and a first test task. Work through this list once after onboarding. Each step links to its detailed page. ## Setup checklist | # | Step | Where | | - | - | - | | 1 | Finish onboarding and create the workspace. | [Getting started](/user-guide/getting-started) | | 2 | Connect at least one model provider (required). | [Model providers](/user-guide/model-providers) | | 3 | Set the workspace default model and thinking level. | Model providers → **Model policy** | | 4 | Choose your interface language and the agents' reply language. | [Settings and notifications](/user-guide/settings-notifications) | | 5 | Check the plan: seats, parallel runs, included features. | [Billing and plans](/user-guide/billing-plans) | | 6 | Invite members and set their roles. | [Members and permissions](/user-guide/members-permissions) | | 7 | Create a project with its folder and write **PROJECT.md**. | [Projects](/user-guide/projects) | | 8 | Create the agents you need. | [Agent management](/user-guide/agent-management) | | 9 | Write each agent's **AGENTS.md**, **SOUL.md** and **MEMORY.md**. | [Agent context](/user-guide/agent-context) | | 10 | Install shared workspace skills and agent-specific skills. | [Workspace skills](/user-guide/workspace-skills) | | 11 | Connect GitHub or GitLab if the work involves code. | [Connect GitHub and GitLab](/user-guide/git-provider-tokens) | | 12 | Optional: Google Drive sync. | [Google Drive sync](/user-guide/google-drive-sync) | | 13 | Prepare test accounts, test data and dev deployment rights. | PROJECT.md and the task | | 14 | Run one small task end to end and check its results. | See below | ## Where to put rules | Information | Put it in | Applies to | | - | - | - | | Project rules: repository, branches, test commands, environments, Definition of Done, granted permissions | **PROJECT.md** | Every agent and session in the project | | How one agent works: process, checks, reporting | That agent's **AGENTS.md** | That agent, every project | | Tone and communication style | **SOUL.md** | That agent | | Identity and role description | **IDENTITY.md** | That agent | | Long-lived knowledge: terms, customers, decisions, lessons | **MEMORY.md** (the agent also writes it) | That agent | | A process reused by several agents | A **workspace skill** | Every agent | | A process for one agent | An **agent skill** | That agent | | Requirements of one piece of work | The task description, acceptance criteria and Task Files | That task | | Passwords, tokens, keys | None of the above. Use provider connections and [Secrets](/user-guide/secrets). | — | A rule for every agent in every project belongs in a workspace skill (for example `team-rules`). Keep each rule in one place: two versions of a rule confuse agents and make them ask for permission again. ## Choose your first agents Start small: one coordinating agent (a Product Owner) and one or two executing agents for your main kinds of work. Add agents when a new kind of work appears. * Give agents clear names; an agent's role cannot change after creation. * Turn on **Orchestrator** only for the coordinating agent. * Give an agent its own model only when it needs a cheaper or stronger one. * After creating an agent, check that it does not show **Setup failed**. * Each agent runs one task at a time; the plan sets how many agents and parallel runs you have. ## Example PROJECT.md ```markdown theme={null} # — rules for every agent ## Goal and scope - Product: ... | Users: ... | Interface language: ... ## Repository and branches - Repository: , folder: - Branch deploys to . Never push directly to . ## Before handing over (always run) - · · · ## Granted permissions (do not ask again) - May merge into when checks pass. - May deploy to . Never touch . - May download: into the task folder. ## Definition of Done - Merged to the right branch, checks green, deployed to dev, smoke check passed. - Proof in Proofs; results in Deliveries. ## When to ask a person - Only for production, money, secrets, deleting data or contradictory requirements. - Ask with options and a default; otherwise decide by the spec. ``` ## Example AGENTS.md ```markdown theme={null} ## Taking work - Read the task goal, scope, acceptance criteria and PROJECT.md. - If something is unclear, state a reasonable assumption and continue. ## Working - Small changes, checked step by step. Nothing outside the scope. - Run every check listed in PROJECT.md before handing over. ## Reporting - Separate confirmed, pending and risks, with links and evidence. - End with blockers (if any) and the next action. ``` Keep these files short and consistent with PROJECT.md. ## SOUL.md and MEMORY.md **SOUL.md** sets voice and style, for example "Reply in English, briefly; use tables for comparisons; avoid jargon with non-specialists." **MEMORY.md** holds what the agent should remember: product terms, customers, decisions with dates, lessons from incidents and links to key documents. Never store secrets or sensitive personal data there. To make the agent remember something, say so in chat ("Remember: from today main deploys to dev"). Review it from time to time in **Manage contexts**; resetting it erases what the agent learned. ## Skills to install first Browse the AgentDesk managed skills and install the ones every agent needs as workspace skills; agents created later inherit them. Create internal skills for repeated processes, such as team rules, a pre-review checklist, a weekly report template or a dev deployment check. To change an AgentDesk managed skill, clone it to edit. After installing, try a message that uses the skill. ## Run a first test task 1. In a project session, ask the coordinating agent to create a small task with clear acceptance criteria. It starts in Inbox. 2. Read the brief, assign the executing agent and move the task to **Ready**. 3. Follow it in the task's **Current execution**. 4. Check **Deliveries** and **Proofs** before accepting Done. 5. Where it stopped, improve PROJECT.md, AGENTS.md or the skill and run again (see [Waiting and Blocked tasks](/user-guide/waiting-blocked)). ## Check a rule change Open a **new** session and ask something that requires the new rule, without naming the file or the answer. A correct answer that shows the agent read the new file or skill passes. If it fails, fix the source (PROJECT.md, skill or context) rather than reminding the agent in chat. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.