Build your first Loop

Start small. Your first Loop should take one piece of text and turn it into something useful. Once that works, add the next step.

In this tutorial you build a Loop that reads a support request and writes a short internal summary for your team. It uses two Loop AI steps. Both steps only work with text. They need no linked agent, never contact another system, and never send email.

Time: about 15 minutes. You need: owner or admin access to an organization in raia Command.

What you will build

The request you will test with is fictional:

Our invoice export stops at 80%. It started this morning. Three people on our finance team are blocked. We have not tried another browser yet.

Part 1 — Create the Loop

1. Sign in at raiacommand.com and choose the organization you want to work in. Then, in the sidebar, click Loops.
2. Click New Loop. raia Command creates an empty draft and opens it straight away. There is no dialog, and the draft starts Paused.

3. Click the pencil next to Untitled Loop.
4. Type a clear name, such as Support request summary. In What does this Loop achieve?, add a short purpose — for example, "Text only, no emails or external actions."
5. Click Save.

6. Stay on the Steps tab.

Why not Build with AI? Build with AI drafts a Loop from a description. It is useful later. For your first Loop, building by hand shows you how each part works.

Part 2 — Add step 1: extract the facts

1. Click Edit Steps.

2. Click Add Loop AI. The add buttons sit at the end of the step list.

3. In Title, type Extract the facts. Kind is already Loop AI processes the result.

4. In Prompt, paste:

Read the support request in the Run objective. Extract the issue, business impact, and missing information. Use only stated facts. Do not invent a cause or claim the problem is fixed. Return three headings: Issue, Impact, Missing information.

5. In What it can see, choose Just the previous step. For the first step, that is the Run objective.
6. Leave Named results (optional) empty. This Loop does not need them.

Why "use only stated facts"? A Loop AI step may produce details that were not in its input. Tell it what evidence to use, then inspect its output. Telling it not to invent a cause or a fix keeps the summary honest.

Part 3 — Add step 2: write the summary

1. Click Add Loop AI again. The new step appears below step 1.
2. Name it Write the internal summary.
3. In Prompt, paste:

Use the facts from the previous step to write a short internal support summary. Include the issue, who is affected, and one sensible next action. Keep it under 120 words. Clearly label missing information. Do not invent a fix. Do not send a message or change any system.

4. In What it can see, choose Just the previous step.

5. Click Save steps.

Saving stores the draft. It does not publish anything, and nothing runs yet.

Part 4 — Set small safety limits

Limits stop a Run that goes further than you expect. This Loop is small, so its limits can be small too.

1. Click the Safety tab.
2. Enter:

• Max steps per run: 4
• Token cap per run: 10000
• Deadline (seconds): 300
• Max external actions per run: 1 (not 0)

3. Click Save limits.

Why 1 external action, if this Loop has none? The cap is a ceiling, not a plan. Setting it to 1 does not add an action, and neither step will ever use it. Do not set it to 0: today, a limit of 0 stops the Run before step 1, even for a text-only Loop. See Runs and troubleshooting.

Part 5 — Describe what "done" means (optional)

1. Click the Outcome tab.
2. Leave Results that must be present empty — this Loop returns no named results.
3. In In your own words (optional), paste:

The final summary uses only facts from this Run's objective. It describes the issue and any stated impact, clearly labels missing information, gives one sensible next action, stays under 120 words, and does not invent a cause or claim a fix.

4. Click Save definition of done.

This wording works for any request, so you can change the test input without editing the outcome. It does not check the specific facts of each sample — you compare those yourself in the Journey (Part 7).

Part 6 — Test the Loop

1. Click Run test in the Loop header.

The window that opens says it uses the draft you are editing, not the published version.

2. In Prompt to Trigger the Loop, paste the sample request. This becomes the Run objective.

3. Check the safety limits shown in the window match what you saved.
4. Click Start run.

Run test really runs your draft. This example's two Loop AI steps do not send messages or change external systems. A test Run still uses tokens. In other Loops, Run test can perform configured external actions. Run a few focused tests.

Part 7 — Check the result

Open the test Run and look at the Journey. It shows each step in order. Click a step to show its full output.

Step 1 should find these facts:

• Issue: the invoice export stops at 80%.
• Impact: three people on the finance team are blocked. It started this morning.
• Missing information: for example, which browser and file format they use, whether there is an error message, and whether it fails for everyone. The request also says they have not tried another browser.

Step 2 should give a short summary that:

1. states the issue, when it started, and who is blocked,
2. says "They have not tried another browser yet." — and nothing broader about what was tried,
3. labels missing information clearly,
4. gives exactly one next action — such as asking the customer to try another browser and share any error message,
5. does not claim a cause or a fix, and stays under 120 words.

Your wording will differ from run to run. That is fine. If step 2 adds anything that is not in the request, tighten the prompt and test again.

In the Run summary, check that both steps finished, usage stayed within your limits, and External actions taken is 0.

Then check the outcome verdict. It is judged by AI and may not pass every Run — treat it as a second opinion, not a status change. If it says text is missing or cut off, expand the step in the Journey and compare its full output before changing anything. If it still disagrees with your own reading, you can simplify or remove the outcome check.

What not to do yet

• Do not add triggers. Keep starting Runs by hand until the Loop behaves well with several different requests.
• Do not add steps that send email or create records. You will learn how to control those in Approvals and external actions.

Try it with more data

Test with two or three more requests before you trust the Loop. Use one that is very short, and one that contains no clear impact:

Login is slow.

Hi, we were charged twice this month for the same plan. Can someone look at it? Thanks, the team at a small bakery.

For the first, step 1 should list almost everything as missing information. That is the correct result.