AI Best Practices for the Best Results

awork AI helps you with individual tasks and recurring workflows. The key is choosing the right AI mode, providing enough context, and reviewing the results. These best practices will help you get more out of AI and spot opportunities to standardize or automate your workflows.

---

1. Choose the right AI mode

Below the chat input, you can choose from three predefined model categories. Start with Balanced and switch only if your task calls for a different mode.

The modes differ in speed, depth of processing, and credit usage. If needed, you can select a specific model, a different reasoning level, or an image model under Erweitert. Which models are available behind the presets is determined by authorized users and admins in the workspace settings.

---

2. Give the agent the right context

The better the agent understands the task and its constraints, the more helpful its results can be.

• Include relevant content: Link relevant projects, tasks, and docs using @, or start the agent directly from a task, project, or comment. It will then use the context of that work, but it won't be added to the task as the assignee.

• Store knowledge in one place: Keep frequently needed briefs, guidelines, and background information in the relevant projects or spaces, for example in awork Docs or centrally in Workspace Kontext. That way, you don't have to provide the same context again and again.

• Check for missing information: Before sending your request, ask yourself whether the agent has all the information it needs to complete the task.

---

3. Keep threads and context organized

A thread is a conversation with an agent. In long threads, older content is compressed, which can lead to details being lost.

These habits will help you stay organized:

• One topic per thread: Whenever possible, use a separate thread for each task.

• Summarize when switching: Before switching, ask the agent to summarize the current status, key decisions, and open questions. Copy that summary into the new thread.

• Link threads: Assign conversations to the relevant project or task so your team can find them and continue the work.

• Use personal memories: You can explicitly ask the awork Agent to remember certain preferences or facts, such as that you prefer answers in bullet points. Personal memories are visible only to you. Custom agents don't currently have a memory feature.

---

4. Improve AI results step by step

You don't have to phrase every request perfectly from the start. Instead, work through it in steps:

1. Describe the task: Explain what you want to achieve and provide the necessary context.

2. Ask for a draft: For bigger tasks, start by asking for a draft or summary.

3. Review the result: Check important facts against the source and look for missing information.

4. Make targeted revisions: Give specific follow-up instructions, for example, to shorten an answer, make the tone more formal, or add dates.

In particular, check numbers, names, and dates before sharing results with customers. For high-impact tasks, don't treat AI responses as final results.

---

5. Choose between a prompt, skill, and agent

If you make the same request repeatedly, you don't need to prompt the instructions from scratch each time. And not every task needs its own configuration. Decide based on how often you do the task and how standardized you want it to be.

A skill describes how to complete a task and can be reused by multiple agents. A custom agent bundles the configuration for a specific job and can be shared with your team.

Examples of recurring tasks

• Format reports consistently: Add a template and an example as supporting materials in a skill, then link it to an agent. This helps ensure that results follow the same guidelines.

• Run weekly project checks: Set an agent to run on a schedule, for example, every Monday morning.

• Automatically respond to changes: Use an agent as an action in an awork automation, for example, to review a brief when a task moves to the "In Review" status.

Automated agents can run on a schedule, in response to a comment, or as part of an awork automation.

---

6. How can I improve my AI agent step by step?

Great agents rarely come from a single perfect prompt. It's only in day-to-day work that you discover which instructions are missing, which edge cases need to be covered, and where the results still don't meet your expectations.

Here's how to improve your agent step by step:

1. Design: Describe the task, the desired outcome, and any important rules.

2. Use: Try the agent on real tasks.

3. Review the results: Check what works and where errors, ambiguities, or missing edge cases come up.

4. Improve the instructions: Give specific feedback. Add recurring corrections to the agent's instructions, its linked skill, or supporting templates so they're taken into account in future runs too.

5. Test again: Use the updated agent on more tasks and check whether the results have improved.

Plan for several rounds of improvements. Results often become much more useful after three to four rounds, though complex tasks may take longer.

---

7. Connect awork AI to your workflows

With connectors, skills, agents, schedules, and automations, you can standardize tasks and integrate other tools.

• Tools in the connector catalog: Install connectors for your tools and add the connector to your agent.

• Tools outside the catalog: Use a custom MCP connector or give a skill credentials for the tool's API.

• Use awork from ChatGPT or Claude: Connect the awork MCP Server to your AI chat. It only accesses content your awork account is authorized to view.

Each connector connection uses one signed-in account, and a connector can have multiple connections. Confidential credentials you provide to a skill remain hidden from other users.

---

8. Protect your data and permissions

awork AI runs within awork. Your workspace data stays private, and awork AI is GDPR-compliant. Neither the agent nor the model learns from your chats. Your account permissions also apply to agents and the awork MCP Server: they can only see content your account has access to.

---

9. Things to keep in mind

What you should do

• Review results before sharing them with customers, and verify important details against the source.

• Have each person set up their own connector connection using their own account.

• For personal credentials in skills, select the Jeder User fügt eigene hinzu setting.

• Share skills and agents with the teams that need them.

What you should avoid

• Don't share personal connector logins or API keys with others.

• Don't put passwords or tokens in a chat.

• Don't treat AI responses as final results for high-impact tasks.

• Don't use one endless thread for every topic.

• Don't switch models in the middle of a thread, to avoid using extra credits.

---

Key takeaways

• Start simple: Use the Balanced model for most tasks and switch modes only when needed.

• Provide context: Include relevant projects, tasks, and documents, and save recurring information in awork Docs.

• Work in a structured way: Keep threads organized, and bring a summary with you when you switch.

• Improve step by step: Ask the agent for drafts, check the facts, and give specific follow-up instructions.

• Standardize repetitive tasks: Use skills, custom agents, and automations when the task calls for them.

• Use integrations thoughtfully: Connect the tools you need through connectors or MCP, and keep permissions and credential handling in mind.

• Stay alert: Check important results and don't share confidential credentials.

This keeps awork AI flexible for one-off tasks while also helping you make recurring workflows more efficient and reliable.

---

FAQ

Which mode should I use?

Use Fast for simple, quick tasks, Balanced for most day-to-day work, and Deep thinking for complex planning and in-depth analysis.

When is it worth creating your own agent?

When you need a dedicated role for a specific job, with its own instructions, model, skills, and connectors. If you want the agent to work without you starting it manually, you can automate it.

Does an agent have access to all workspace data?

No. Agents and the awork MCP Server can only see content the relevant awork account has access to.

Does awork AI learn from my chats?

No. Neither the agent nor the model automatically learns from your chats. However, the awork Agent can save personal preferences and facts as memory across threads and take them into account in the future. These memories are private and tied to your user account. Custom specialized agents don't currently have an automatic memory feature. Instead, you can add proven instructions to their prompt or linked skills so they're taken into account in future tasks too.