> ## Documentation Index > Fetch the complete documentation index at: https://dataleap-469bb19d-docs-from-request-2026-08-21-templates.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Your first agent > Create an agent from a plain-language request and make it reliable through feedback. One chat window is one agent. You describe the job, Dataleap builds it — and everything after that happens in the same chat: refining, connecting tools, giving feedback. ## Build from a prompt Open **New Agent** and describe the outcome you want. If you know exactly what you need, be precise. If not, use the most powerful prompt there is: > "Ask me questions so you understand thoroughly what I need." The angle is harder than the writing — answering questions is easy, writing specs is hard. Attach supporting files if you have them. Dataleap asks about inputs, outputs, and edge cases. Answer them all — this is where the agent gets the context it needs to build exactly what you want. Dataleap tells you which tools the workflow needs and asks you to connect them — you just confirm. Run the agent and check the output against what you expected. Once the test output looks right, enable the trigger and let the agent work. You'll keep refining it from real results. ## Write better first prompts * Start with one job, not a list of unrelated tasks — one chat window is one agent. * Include the source of truth the agent should use. * State the final output you want. * Add examples if the format matters. ## Show, don't just tell The richer the context you give, the better the output. Instead of describing what's wrong ("the output looks off"), show it: attach a screenshot and say exactly what to change — "make the headers actual bold, include the emojis in the text, structure the formatting so it's readable." Share screenshots, reference posts, style guides, voice notes, images, links. **Context is free — use it generously.** ## Improve it from real runs Once the agent is running, feedback happens in the same chat: when a result isn't what you wanted, tell the agent specifically what was wrong and what it should have been. You don't need to dig through settings — just prompt it. **Train it, don't judge it.** The first version might disappoint you — that's normal. A new coworker needs time to learn your preferences; your agent is no different. Every correction you give is saved permanently and compounds on every run after: the agent that struggled last week can be excellent this week because of your input. And unlike your best team members, it never forgets a rule and has never once called in sick. See the build flow in action. Next: make the behavior predictable and safe. How connections work when Dataleap asks for them.