Best practices

This page collects practical, end-user guidance distilled from the full documentation. Bookmark it as a quick reference once you have worked through the individual feature guides.

Writing effective prompts

The quality of the model's output is directly tied to the clarity of your input. A few habits make a significant difference:

• State your goal upfront. Begin with what you want, not with background context. The model reads the whole message, but leading with the objective helps it weight the rest correctly.
• Specify format and length. If you need a bullet list, a table, or a one-paragraph summary, say so. Left unspecified, the model makes its own choice, which may not suit your use case.
• Give examples when the task is ambiguous. Showing one or two input-output examples (few-shot prompting) is the fastest way to convey subtle requirements — tone, style, schema — without lengthy instructions.
• Assign a role where it helps. Prefixing with "You are an expert in…" or "Act as a senior editor…" primes the model for domain-appropriate vocabulary and reasoning depth.
• Iterate rather than over-engineer. Start simple. If the first response misses the mark, refine the prompt incrementally rather than writing a very long instruction from scratch.

Save prompts you use repeatedly as templates in the prompt library so you do not need to retype them. See the Prompts guide.

Choosing the right model for a task

Not every task needs the most powerful (and most expensive) model. Matching capability to requirement saves cost and often produces faster responses:

• Simple Q&A, classification, or extraction — use a fast, economical model such as GPT-4o mini, Claude 3 Haiku, or Gemini Flash. These handle straightforward tasks quickly and count fewer messages against your monthly cap.
• Complex reasoning, long-document analysis, or code generation — use a flagship model (GPT-4o, Claude 3.5 Sonnet / Opus, Gemini Pro). The higher capability is worth the additional cost for tasks where accuracy matters.
• Creative writing and open-ended generation — different providers have different stylistic strengths. Experiment across two or three models on the same prompt to find the output style that fits your needs.
• Long contexts — if your task involves large files or long chat histories, check the context-window size on the Models page before selecting a model. Choosing a model with an insufficient context window will cause earlier messages or file content to be truncated.

Using tools effectively

Tools extend what the model can do — web search, file reading, code execution, and more. A few principles make tool use more reliable:

• Enable only the tools relevant to your task. A model with fewer tool options makes cleaner decisions. Unnecessary tools add noise and can lead to unintended side effects.
• Be explicit when you want a tool used. Saying "search the web for current prices" is clearer than "what are current prices?" — the former signals that real-time lookup is required.
• Review tool results before acting on them. The model summarises tool output; for critical decisions, expand the tool-call trace in the chat to inspect the raw result.
• Chain tools intentionally in workflows. If you need multiple tool calls in sequence (search → fetch → parse), use a workflow rather than asking the model to improvise the chain in a single conversation. Workflows are more predictable and auditable. See the Workflows guide.

For the full list of available tools and how to configure them, see the Tools guide.

Getting the most from agents

Agents are reusable, pre-configured AI assistants. They save time on recurring tasks and ensure consistent behaviour across team members:

• Write precise system prompts. A vague system prompt produces inconsistent output. State the agent's role, the tasks it should and should not perform, the preferred response format, and any domain rules.
• Test before sharing with the team. Run several representative queries against a new agent in a private conversation before making it available to teammates. Fix edge-case behaviour in the system prompt rather than after the fact.
• Use one agent per job. Narrow, focused agents outperform broad, do-everything agents. If you find yourself writing a very long system prompt covering unrelated tasks, split it into two agents.
• Combine agents in workflows for multi-step tasks. Orchestrating several specialised agents through a workflow is more maintainable than a single monolithic agent trying to do everything.

See the Agents guide for configuration details.

Organising your chats

As your chat history grows, a little organisation goes a long way:

• Rename chats immediately after starting them. The auto-generated title is a guess based on your first message. Renaming to something descriptive (e.g. "Q3 campaign copy — first draft") makes it easy to find later.
• Start a new chat for each distinct task. Mixing unrelated topics in one conversation confuses context and makes the history harder to search.
• Archive completed chats. Archiving keeps your main list focused without permanently deleting the conversation — you can search and restore archived chats at any time.
• Use team chats for collaborative work. If you are working in a team workspace, starting the conversation there means teammates can continue or reference it without needing a separate share link.

See the Chats guide for chat management features.

Managing cost and usage

Your monthly plan cap covers all messages sent across every chat, agent, and workflow. These habits help you stay within your allowance:

• Use economical models for exploratory work. When you are drafting, iterating, or experimenting, reach for a smaller model. Switch to a more capable model only when you are finalising.
• Avoid attaching large files unnecessarily. File content is included in the message sent to the model and counts toward your usage. Attach files only when the model genuinely needs the full content.
• Review workflow step counts. A workflow that calls five agents in sequence generates five messages per run. For high-frequency workflows, consider whether any steps can be consolidated or replaced with a deterministic tool step.
• Monitor your usage in Settings. Check Settings → Usage to see how much of your monthly cap you have consumed and when it resets. Enable usage alerts in your notification preferences so you are warned before reaching the limit.
• Upgrade early if you are a heavy user. Upgrading mid-month restores your full cap immediately rather than waiting for the next billing cycle.